Total addressable market—TAM—is one of the first analytical exercises any B2B business undertakes, prompted by the question: how large is the universe of potential revenue available if the business could sell to everyone who might plausibly want what the business has to offer? This estimate is almost always computed top-down, starting with the team’s opportunity-sizing of the overall market, then applying filters for geography, segment, and product scope to arrive at the relevant slice. The result is a single large number: a quality check for early strategic decisions and a framing device for investors who want to know if the market is large enough to even justify this business.
But TAM in this form is a monolithic construct. It produces one number, or at best a few numbers by segment, and it says nothing about the individual accounts (customers and prospects) that collectively make up that market. When we see a statistic like this, say $90 billion total addressable market in North and South America, we might want to know which individual accounts make up that $90 billion. Or, how the $90 billion is split across the accounts. Or which industries have the deepest proven adoption. Or, maybe, which territories have the most balanced potential across customers and prospects. Perhaps if we can split the TAM by sales reps and account managers, and if so, how we would go about that.
These were the questions I had to answer in my work with go-to-market (GTM) teams who needed a way to allocate customers and prospects across sales reps and account managers in a mathematically rigorous, optimizable way. To build a GTM-centric optimization model for territory assignments, I needed a single, comparable measure of revenue potential for every account out there. Not a rough estimate based on firmographics, not an arbitrary score, but a number that meant the same thing across industries and company sizes—a number grounded in actual observed behavior. That number, as it turned out, was TAM, just a different flavor of it; one built from scratch, account by account, using the customer behavioral data accumulated from years of operating the business. I called it the Modular TAM.
What is Modular TAM?
Modular TAM is a framework for computing a defensible revenue potential for every account in the market—customers and prospects alike—using benchmarks derived from the company’s own customer data. It is calibrated specifically to each company’s product, observed penetration patterns, and historical pricing, which is precisely what makes it defensible and precisely why it requires real data to build.
It’s modular because of its structure: four independent components, each answering a distinct question about the account, each calibrated separately. We can update the penetration benchmark without touching the pricing benchmark, swap the data cut from one feature to another without rebuilding the formula, and so on. The model evolves as the data grows.
Each account gets its own TAM value, grounded in the behavior of comparable accounts in the customer base, and those values can be rolled up however needed—by industry, region, territory, customer versus prospect—to produce a bottom-up market view where every dollar traces back to a specific account and a specific set of benchmarked assumptions. That traceability is what makes the framework useful for territory assignment optimization: when every account carries a comparable, modeled dollar value, territory assignment becomes a proper optimization problem with real inputs.
What we need to build this
The framework requires four categories of data for the company’s customers and prospects: firmographics (employee count, industry, revenue, and so on, available from vendors like ZoomInfo or from public filings); product usage data, specifically active users per account from the company’s product database; financial data, meaning revenue per account from the billing system; and growth rates, from data vendors or historical snapshots.
The quality and scope of the customer and prospect data matters enormously. Two to three years of customer history across a few dozen accounts in multiple industries is the practical minimum for computing benchmarks that are worth trusting. The more accounts, the more years, and the more industry diversity in that base, the tighter and more defensible the benchmarks become. Notice the implication—this is not a methodology to use before that foundation exists. But once it does, the methodology is fully computable from data that most mature B2B businesses already have sitting in their systems.
Modular TAM formula
The core equation has four components, multiplied together:
Each component answers a different question about the account. Addressable users answers how many people at this account could realistically use the company’s product. Penetration rate, expressed as percentage, answers what fraction of those addressable users will actually be provisioned, benchmarked by industry (or other meaningful feature) from the company’s customer data. Revenue per user per year answers what the company will earn annually per provisioned user, also benchmarked by industry (or other meaningful feature). And Projected years (T) sets the time horizon over which we want to size the opportunity.
Addressable users: who can use the company’s product?
If the company’s product is something that anyone in any role could use—productivity software, payroll tools, company-wide learning platform—then for addressable users we can simply use the account’s total employee count from the firmographics data. If it serves a more specific function, like a CRM for salespeople or an IDE for engineers, then addressable users is that team’s headcount, which we either source from a data vendor or estimate as a percentage of total employees based on industry norms.
The growth projection is compound: today’s addressable users multiplied by (1 + annual growth rate) raised to the power of T, where the growth rate is typically proxied by headcount growth, publicly available for large companies and estimable from data vendors for smaller ones.
This is the one component of the framework that does not require customer data. We can compute addressable users for a prospect the company has never sold to just as easily as for an existing customer. The next two components are fundamentally different—they require a benchmarking step that starts from the company’s own customer base, which is exactly why the two-to-three year data prerequisite matters.
Benchmarks and the generalized mean
Before turning to penetration and pricing individually, it is worth understanding the mathematical tool that powers both of them, because both benchmarks use the same formula—the generalized mean—just with different values of a single parameter.
The generalized mean is a family of averages, controlled by an exponent p:
When p = 1, we get the arithmetic mean, the familiar average that treats all observations equally. When p = 3, which we can call the cubic mean, the formula raises each value to the third power before averaging, then takes the cube root of the result. This amplifies larger values before the average is computed, producing a benchmark that leans toward the high end of observed values. When p = ⅓, the cube root mean, the opposite happens: smaller values get amplified, and the benchmark leans toward the conservative end of the distribution.
The insight that makes this useful here is that the choice of p encodes a strategic stance directly into the mathematics. For penetration, we want to be ambitious. The question the benchmark should answer is: given what we have actually observed in accounts comparable to this one, how deep could penetration realistically go? The cubic mean (p = 3) answers this by giving more weight to the accounts in our customer base where deep adoption has already been proven. It does not extrapolate beyond observed data; it simply rewards the signal that already exists.
For pricing, we want to be conservative. The question is different: not what have our best accounts ever paid, but what will this account actually pay? The cube root mean (p = ⅓) pulls the benchmark toward the lower end of observed pricing, protecting the model from being inflated by a handful of enterprise accounts that pay premium rates most accounts will never match.
Ambitious on penetration. Conservative on pricing. Same formula, different exponent.
Penetration rate: how deep could this customer or prospect go?
The benchmarking process for penetration has two steps. First, we compute the current penetration rate for each existing customer account: all-time active users divided by addressable users today, multiplied by 100. This gives us one percentage per account—a concrete, observed measure of how deeply the company’s product has been adopted.
Second, we benchmark by the data cut that best differentiates penetration rates across our customer base. We test candidate attributes—industry, region, ownership, business model, and so on— and select the one along which penetration varies most. If healthcare accounts consistently penetrate at 18% and technology accounts at 6%, industry is doing real explanatory work and becomes our cut. Within each group, we compute the cubic mean of the individual account penetration rates, arriving at one benchmark per group, a number that leans toward the accounts where deep adoption has been proven, but remains entirely grounded in observed behavior.
When we encounter any account, be it a prospect the company has never sold to or an existing customer we are sizing for expansion, we look up their group, retrieve the benchmark, and plug it in. The benchmark answers a specific question: given accounts like this one in our customer base, how deep could penetration realistically go?
Revenue per user per year: what will they actually pay?
The pricing benchmark follows the same structure, with the opposite exponent.
First, we calculate revenue per user for each customer account. There are two reasonable approaches: the all-time view, which divides total revenue ever earned from the account by total users ever active, and the trailing-12-month view, which uses recent revenue and recent active users. The all-time view smooths out fluctuations; the trailing view reflects current pricing more accurately. Either works; the right choice depends on how much the company’s pricing has evolved over the period in question.
Second, we benchmark via cube root mean (p = ⅓) using the same data cut we identified for penetration. Within each group, the cube root mean of the per-user revenue values pulls the benchmark toward the lower end of observed pricing, creating a conservative counterpart to the ambitious penetration benchmark.
Bounding the TAM: ceiling and floor
Even with carefully benchmarked inputs, the unbounded formula can produce numbers that are not grounded in financial reality. A Fortune 500 company with 200,000 employees might produce an Unbounded TAM of $50 million—technically consistent with the math—but that number is meaningless if the buying department’s entire annual budget is $3 million.
The ceiling is derived from the account’s financials, working downward in two steps. We estimate the relevant department’s budget as a percentage of total company revenue, and then estimate the product category’s share (of our company) of that department budget. Multiplying both percentages by the account’s cumulative revenue over T years gives us the maximum the account would realistically spend on a product like ours. The cumulative revenue calculation matters here—a company generating $1 billion today at 10% annual growth produces substantially more cumulative revenue over five years than a flat extrapolation of 5 × $1 billion would suggest.
The floor is simpler: β multiplied by the account’s current annual revenue, where a reasonable β value is 2. If the model produces a TAM lower than what an account is already paying us, something has gone wrong in the inputs and the floor catches it. Notice that it works for prospect accounts as well; because their annual revenue is $0 (the company hasn’t signed them yet), the floor is also $0.
The final TAM for each account is therefore:
Final TAM = max ( Floor, min ( Unbounded TAM, Ceiling ) )
Bounded below by what we have already proven the account will spend, and above by what the account can realistically afford.
What Modular Total Addressable Market enables
With a modeled dollar value attached to every account, from customers to prospects, several things become possible that a market-level TAM cannot support.
Account prioritization becomes data-driven rather than instinct-driven: we can rank prospects by Modular TAM and direct sales development effort toward the accounts with the largest modeled opportunity, not just the most recognizable logos.
Whitespace analysis, for existing customers, becomes concrete. The gap between an account’s current annual revenue and its Modular TAM is the expansion opportunity, and that gap is invisible without this framework. An account paying $200K with a TAM of $2M has ten times the whitespace of one paying $200K with a TAM of $300K, but in a standard GTM analysis, those two accounts are completely indistinguishable.
Territory design, as noted at the outset, is where Modular TAM was built to operate. When every account has a comparable dollar value, territory assignment becomes an optimization problem in the formal sense: balance territories by total TAM, ensure each sales rep or account manager carries a comparable opportunity—or place more skilled people against higher-TAM books of business—while controlling for geography and segment. That optimization is tractable precisely because the inputs are consistent and at account-level.
And forecast calibration becomes possible in a new way: by comparing pipeline projections against TAM-derived ceilings, we can surface cases where a forecast exceeds what the model says is plausible for a given account and prompt the right conversations before it’s too late.
Every year, the same ritual plays out inside B2B go-to-market organizations. Sales leadership gathers and they forecast the coming fiscal year’s revenue. The process is straightforward: each account manager looks at their book of business, mentally assesses the health of each deal, and assigns a probability that the deal will close—often, though not always, based on which stage it is in. The numbers then trickle upward to region-level views and eventually land on the desk of the Chief Revenue Officer as a single, confident figure.
These forecasts are not insincere. Account managers have context that no dataset can replicate. They’ve sat across from the buyer, they’ve read the room, they know that the VP of Procurement at one account just got promoted or that another account’s finance team has been ghosting them since October. This intelligence is invaluable.
But when an account manager (or, any of us for that matter) says a deal will close at some probability, they are making a point estimate—a single scenario in which the world behaves exactly as they expect. When every account manager does this simultaneously, and the numbers are summed, the result is a forecast that assumes every account manager’s best guess is correct at the same time. This almost never happens. Some will be right, others will overshoot, and a few will be blindsided by churn no one saw coming. The forecast, in other words, is not really a forecast. It’s a single scenario.
This is the gap I set out to close during my time working with go-to-market executives. What follows is a framework for bridging the distance between a believable forecast and a probable one. I call it Two-Lens Forecasting. The core idea is simple: instead of choosing between the team’s human forecast and a data-driven model, we use both—one lens from forward-looking, on-the-ground intelligence and one from historical, probabilistic intelligence. Neither lens alone gives us the full picture, but together, they lock in a more precise target.
The individual methods we’ll talk about are not new. Monte Carlo simulation has been used in finance and engineering for decades. Beta distributions are a staple of Bayesian statistics. Weighted-pipeline forecasting is standard practice in most go-to-market organizations. What I believe is valuable is the specific combination: a dual-methodology simulation engine (one approach for renewals, a different one for new business and expansions) paired with a visual deep-dive framework (which we will call “pill charts”) that makes the output immediately useful to executives.
Lens one: The weighted-pipeline forecast
To make this concrete, consider a B2B company doing a revenue forecast for the coming fiscal year. The CRO has been given these numbers—$200 million of revenue is in active pipeline and the team believes they can land $155 million: $80 million from renewals, $38 million from new business, and $37 million from expansions. These totals are based on region-level forecasts derived with a weighted-pipeline approach:
Overall, the forecast’s message is this: there is $200 million of revenue that could be locked in and the team thinks they can close 78% of it. Sounds reasonable. But what if North America renewals all slip at the same time? That’s $30 million gone down the drain. What if deals in Europe, Asia, and South America are all won but all deals in North America are lost? That’s a total of $100 million won, a far cry from $155 million. Our gut instinct should tell us that such extremes are less likely to happen but that they aren’t impossible. It should also make us wonder about the reverse: $155 million is possible but how likely is this possibility? This is where weighted-pipeline forecast falls short and where we need to use a Monte Carlo simulation.
Lens two: Monte Carlo simulation
The core idea of Monte Carlo is disarmingly simple: instead of computing a single expected outcome, we simulate thousands of possible outcomes and look at the distribution of results.
Say we run this simulation across the entire $200 million pipeline—renewals, new business, and expansions combined—and get the following distribution:
The model’s most likely outcome is $107 million, not $155 million. The 80% confidence interval runs from $85 million (downside—things went badly but not catastrophically) to $130 million (upside—nearly everything went our way). The team’s $155 million sits in the top 5% of simulated outcomes. That doesn’t mean it’s impossible but we now know something we didn’t with the weighted-pipeline approach: $155 million requires strong performance across renewals and new business and expansions simultaneously. Barring exceptional luck, that outcome is unlikely. We need to pressure-test what’s making the (human) weighted-pipeline forecast so optimistic.
How the simulation works: Renewals
Before we do that, let’s first address how we set up the Monte Carlo simulation correctly. We need to treat the three deal types—renewals, new business, and expansions—differently because they behave differently.
Renewals have a concrete plan base: the dollar amount up for renewal. The question is what percentage of that plan base the customer will actually renew. The most honest approach is to let history speak directly.
Say we pull 1,800+ historical renewal opportunities across two fiscal years and plot the distribution of renewal win rates. What we’d likely see is striking and bimodal: most deals either churned completely (0% of plan dollars renewed) or renewed fully (90–100% of plan dollars renewed), with a thin scattering of partial renewals in between. This becomes the empirical distribution Monte Carlo draws from: the actual, observed renewal rates from history.
But not all renewals are the same. Let’s say we test several stratification schemes and find that two variables emerge as the strongest predictors of renewal rates: region and deal size. Region sorted into three tiers—high-win, medium-win, and low-win—based on historical aggregate renewal rates. Deal size sorted into three buckets: small (under $20K), medium ($20K–$500K), and large ($500K+). Crossing these two dimensions produces a 3×3 grid of nine profiles, each with its own empirical distribution (histograms) of renewal rates.
For each renewal deal in the current fiscal year’s pipeline, the simulation proceeds as follows:
Identify the deal’s profile (its region tier × deal size tier).
Sample a renewal rate at random from the empirical distribution of historical renewal rates for that profile.
Multiply that sampled rate by the deal’s plan renewal amount to get the simulated renewal revenue.
Repeat for every deal in the portfolio.
Sum across all deals for one simulated total.
Repeat the simulation ten thousand times.
How the simulation works: New Business and Expansions
Next come the new business and expansion deals. These usually don’t have a plan base to renew against. The relevant question is not “what percentage of an existing contract will be retained?” but “will this deal close, and if so, at what amount?”
This is where stage-based conversion rates come back into the picture, but with a probabilistic upgrade. Instead of taking the historical conversion rate as the only option, we can model each deal’s close probability using a Beta distribution, parameterized by the observed successes and failures at each stage.
Let’s say one of the deals we’re interested in is currently in early-pipeline stage, associated with a 40% win rate. The question we should be asking ourselves is 40% of how many historical deals? 40 out of 100? Or 4 out of 10? Intuitively, we would have more confidence in the former.
The Beta distribution helps us encode this. If we’ve observed 40 wins and 60 losses at this stage historically, the Beta distribution would say something along the lines of, “The conversion rate is probably around 40%, but given our sample size, it could plausibly be anywhere from 32% to 48%.” If we’ve observed 4 wins and 6 losses, the Beta distribution would widen this interval: the conversion rate could be anywhere from 15% to 68%.
For each new business or expansion deal, the simulation works as follows:
Identify the deal’s current stage.
Sample a win probability from the Beta distribution associated with that stage.
Simulate a binary outcome: the deal either closes (with probability = sampled rate) or doesn’t.
If it closes, use the deal’s forecasted amount as the revenue contribution.
Repeat for every deal.
Sum across all deals for one simulated total.
Repeat ten thousand times.
This method captures something the weighted-pipeline revenue forecast completely misses: a portfolio of early-stage deals with thin historical data is fundamentally less predictable than a portfolio of late-stage deals with deep historical data, even if their weighted-pipeline values are identical. The Beta distribution’s sensitivity to sample size makes this explicit.
Two-lens forecasting: The pill chart
Let’s go back to pressure-testing. Monte Carlo told us the most likely outcome is $107 million of revenue closed, and specifically that there is 80% probability the actual value will land somewhere between $85 million and $130 million. The team, on the other hand, said they think they can close $155 million. Something is off but we still don’t know why.
This is where we can visualize each forecasting lens via a “pill chart,” broken down by whichever data cut represents how the business runs its sales engine—in our case, by region.
When the vertical line (weighted-pipeline forecast) falls inside the pill (Monte Carlo forecast), the two approaches are aligned—the team’s projection is within the range that historical data suggests is plausible. When it falls outside the pill, a conversation is needed: what does the team know that the model doesn’t? Where exactly is the team’s assumption diverging from the model’s?
This framing is essential. The point is never to tell a sales leader that their forecast is wrong. The model’s output is the rearview mirror; telling us what would happen if the future looked like the past. The team’s forecast is the windshield; informed by on-the-ground intelligence about buyers, deals, and market conditions.
In our case, Europe and Asia look good: the past and the future align. But North America warrants further investigation and South America looks unreasonably optimistic—the model’s best-case scenario is $16 million and the weighted-pipeline forecast is $30 million. To understand this, we double-click into the three deal types for these two regions and make another pill chart:
This decomposition surfaces insights that neither approach can produce alone. In North America, the renewals are on track; the team’s $30 million sits comfortably inside the model’s range. The gap is cleanly isolated to new business and expansions, where the team seems to have conviction that net new revenue (from new logos and expanded customers) will convert at higher rates than it has historically. Go-to-market executives now know they need to talk to account managers who cover North America and pressure-test their assumptions—are they certain of their conviction? Why? If they can explain, great. If not, we need to scale back the forecast.
South America tells a different and more concerning story. Three problems at once: the team’s renewals forecast of $14 million is nearly double the model’s best-case of $8 million. The new business forecast of $5 million is actually below what the model suggests is likely—the team may be leaving money on the table. And the expansions forecast of $11 million exceeds the model’s range of $5 million to $8 million. The executives now know the South America team needs to go back to the drawing board; either the team is seeing a profound behavioral shift in this region or the on-the-ground intelligence is scattered. If the former, the team might need to rethink the entire sales motion in this part of the world. If the latter, the executives might need to have tough conversations with the team.
Either way, the pill charts tell a clear story:
Unless we have defensible conviction that new business and expansions deals will perform better in North America than they have historically and that a profound behavioral shift is happening in the South America market, the $155 million target is unattainable.
If we have evidence in one but not the other, then a more realistic yet still ambitious target is closer to the 75th percentile of the Monte Carlo distribution—around $130 million.
And if we have no defensible conviction for either—in other words, no evidence that future will be different from the past—it’s best to stick to $107 million (the median) as the target.
Limitations
It would be a disservice to present this framework without addressing its blind spots. It has two, and naming them matters—a forecasting tool whose limitations we don’t understand is more dangerous than no tool at all.
The first is correlation. The Monte Carlo simulation treats each deal as an independent coin flip, but in reality, macroeconomic shocks, competitor moves, and budget freezes sweep through entire segments at once. The consequence is that the model’s confidence intervals are likely too narrow. The true tails are wider than the simulation suggests. Bridging the simulation with the weighted-pipeline forecast partially compensates for this: sales leaders who sense that “the whole market is tightening” are implicitly injecting correlation, even if they’d never use that word.
The second is thin cells in the stratification grid. Not all nine cells in the 3×3 matrix have equal historical depth. A corner cell might contain a few dozen data points; enough to sample from, but noisy enough that another year of data could meaningfully shift the distribution. Using the empirical distribution directly (rather than fitting a parametric model) avoids imposing false structure, but the honest answer is that some cells deserve an extra grain of salt.
The philosophical point
The two lenses represent two fundamentally different epistemologies. The human forecast starts from what the team believes will happen and works backward to a number. The Monte Carlo forecast starts from what has happened and works forward to a range. One is rooted in the future, the other in the past. Neither is right in isolation. A forecast built entirely on the past would ignore the intelligence salespeople accumulate through months of relationship-building, while one built entirely on the future would ignore the base rates that tend to reassert themselves. That’s what Two-Lens Forecasting is really about—not choosing between the past and the future, but insisting on both.
I had never witnessed anything like it. The day my friend and I walked from the Taj Mahal Palace to Marine Drive, I was told to follow her lead, for crossing the street in Mumbai was an art only the locals had mastered, but even she was overcome with doubt when we found ourselves at the braided intersection near the Mantralaya. There, an interlock: pedestrians hustling between moving scooters and cars and buses, sealing the gaps between the vehicles like rivulets flowing through cracks in arid soil. When they needed—no, wanted—to stop the traffic and move through, the pedestrians made eye contact with the drivers and held their hands out, turning every interaction into a micro negotiation between life and death. Honking, honking, so much honking: drivers at pedestrians then drivers at each other, a sonic assault spasming to the rhythm of this infrastructural chaos, of moments which appeared like stochastic events, without any precedent or pattern. I stood still, watching it all unfold, and then a thought crossed my mind—that I could not imagine a self-driving car in this place.
In July 2025, the political organization GrowSF ran a pulse poll on autonomous vehicle support. 67% respondents were in support of self-driving cars, and of those supporters, 30%—the largest subset by percentage—supported having such vehicles on San Francisco roads because they were “safer than human driving.” This struck me as both familiar and unusual. I heard the same reasoning whenever I had asked friends why we took Waymos in the city and yet, before self-driving cars became open to the public, I had never heard any of us talk much, if at all, about a desire to have something safer than a human behind the wheel. I had trouble imagining that, by extension of this poll, some 72,760 San Franciscans could be choosing a self-driving car for this reason. That a more intuitive reason—autonomous vehicles being safer than ride-shares—had been cited less frequently (16%) than autonomous vehicle services being dependable and efficient (18%) was also telling. Over the past two years of (enthusiastically) using Waymo in San Francisco, dependability and efficiency were almost never the reasons why I opted for a self-driving car: on occasion, the car didn’t show up, and almost always, the pickup time was three to five times longer compared to a rideshare. I didn’t doubt the poll numbers; they reflected what we told ourselves but I had started to wonder if what we told ourselves was merely a cognitive reframing.
Five years ago, I was visiting my family in Mostar and had to pay a visit to the local government office to submit some paperwork. I asked a friend to be my chaperone because I had lost the muscle needed to navigate the psychological labyrinth that was the bureaucracy in the Balkans. She worked as a legal assistant so she knew the unwritten customs of administrative undertakings: when to stand in line (never), who is allowed to cut you in line (a parent with a small child), which counter to approach to pay the fee (you just know), which counter to approach to submit the paperwork (you just know), which booth to approach to have the photo taken (you just know), when to crack a joke at the counter (you feel it), when to thank profusely at the counter (always). I remember looking around the office for signs, notices, instructions—anything, really—that codified the protocols my friend knew so well, but none were in sight. Something else happened during this visit: my dad told me about a “foreigner” who had recently driven through a one-way street in Mostar in the wrong direction. The people who had witnessed this happen were allegedly aghast. I asked my dad how the guy could have known the right direction; at the time, few if any streets in Mostar had one-way designation signs and, unless you were a local who had grown up and learned to drive in the city, Mostar’s streets read like an esoteric language. It took me a while to understand why he never answered my question. The answer was too obvious: it’s not that the guy should have known the right direction, it was that he should have known to look for a chaperone.
It was quiet. We were standing on the overpass and looking at the traffic below. I did not hear a honk since my friend and I had arrived in Singapore and I was, somehow, not nearly as attuned to the lack of noise as I was to the spaces between the vehicles. They moved in unison: never getting bigger and never getting smaller, at a pace so steady that I was convinced everyone was driving at the same speed. My friend had pointed out at breakfast to her friend—a native Singaporean who was showing us around the city—how orderly the culture seemed, in a way an inverse of India’s culture in which she grew up. Uncertainty was too much for Singaporeans, he said. Time was used to work and build wealth; predictability was essential to optimize for that goal. I got a hunch on the overpass that a traffic of only self-driving cars would likely look like the traffic in Singapore: cars spaced equally, gliding in silence at the same speed. An equilibrium.
The nominal GDP per capita of San Francisco had been $317,946 in 2023. That had been almost four times the value of Singapore’s GDP per capita that year, at $85,412. Singapore’s GDP, in turn, had been six times higher than Mostar’s GDP per capita, at $13,245, which in turn had been almost two times higher than Mumbai’s GDP per capita of $7,700. Which is to say it was the 72,760 people with $317,946 of goods and services produced per head who might have chosen a self-driving car because it was “safer than human driving” and it was the drivers with $85,412 of goods and services per head who drove in unison with an equal amount of space on the road and it was the chaperones with $13,245 of goods and services per head who could have told you which counter to approach and which direction to take on a one-way street and it was the pedestrians with $7,700 of goods and services per head who held their hands out and turned every interaction with drivers into a micro negotiation between life and death.
To live is to negotiate. Life is predicated on small but unending give-and-takes, what we perceive as social friction—the force in our cultural fabric generated by interdependency on one another, the force which opposes avoidance of reciprocity, that which injects a sliver of randomness into each day. It’s what’s there when we have to put in a customized and complicated food order at the counter, when we have to call customer service and gauge how we need to phrase our complaint, when we have to approach someone we’re interested in, when we have to sit in the car with a driver we likely don’t know who might be in the mood to talk or might be listening to music we don’t like or might not be okay with us talking on the phone. It’s what’s not there at the self-ordering kiosk machine, in the chat window with an agent who might or might not be human, in the left or right swipe on a dating app, in the backseat of a self-driving car. Flexibility, speed, ease, connection, safety—the promises on which technological advancements have arrived are seductive anchors, and I in no way mean to generalize that such advancements are pernicious or useless, but I wonder, now more often than before, if they are, at the same time, high-budget extensions of our tendency to reduce the emotional wear-and-tear bound to happen in a life of social friction. And if so, is it even fair to pass such a judgment? If each of us were in a system where a high value of goods and services produced per head could justify investments which reduced social friction, would not many of us give in? And yet, is it unfair to question our innate tendency to seek comfort whenever provided? If each of us lived in a state of persistently high social friction, would we all not come to see that life, however haphazard, could still happen?
References and Calculations
Number of San Franciscans who chose a self-driving car primarily because they believed it was safer than human driving
California Public Utilities Commission reported 1,016,546 Waymo rides in September 2025 across California (CPUC AV Program Quarterly Reporting; archived on Nov 24, 2025; monthly-level deployment dataset was downloadable via Reporting Period: July 1, 2025 – September 30, 2025 zip file). At the time of reporting, Waymo was the only carrier reporting the data about the operation of their vehicles in the Commission’s AV Passenger Service Deployment programs. Other carriers existed but were not widely deployed.
In November 2025, Waymo’s driverless deployments in California were fully operational in the San Francisco Bay Area and Los Angeles (Waymo map, archived on Nov 24, 2025); through June 2025, Waymo had reported 29.888M rider-only miles in San Francisco and 16.462M miles in Los Angeles, thus San Francisco accounted for roughly 65% of Waymo’s driverless miles in California (Waymo Safety Impact; archived on Nov 25, 2025). Also in June 2025, TechCrunch reported that Waymo had “600+ vehicles” in San Francisco and “400+ vehicles” in Los Angeles (Korosec, K. ‘Waymo robotaxis are pushing into even more California cities‘; archived on Nov 25, 2025); this could be interpreted that roughly 60% of Waymo’s California fleet was in San Francisco. While these figures (65% of miles in San Francisco, 60% of fleet in San Francisco) are not necessarily proportional to Waymo rides, I used the latter (more conservative) to make an assumption that roughly 60% of the 1,016,546 Waymo rides—about 609,930—in September 2025 across California were in San Francisco.
My own Waymo account history reported in the 2024 end-of-year summary that I was in the top 4% of users with a total of 22 trips. In 2025, through the end of November, I had a total of 67 trips, averaging roughly 6 rides per month. I assumed my behavior in 2025 still reflected the behavior of the top few percent of users (my ride frequency increased simply because Waymos became more common in San Francisco) and that most users in San Francisco averaged fewer monthly rides. Because I didn’t know the exact average number of Waymo rides per San Francisco user, I considered a range of plausible values instead of assuming a single number.
Based on general intuition about how often typical residents use the service, I assumed the average could reasonably fall anywhere between 1.5 and 4 rides per month. Rather than treat each value as equally likely, I assigned rough confidence levels to a few representative points in that range: 10% confidence that the average is as low as 1.5, 40% confidence around 2.5, 30% around 3, and 20% around 4 rides per month. These confidence levels simply reflect my judgment that mid-range values were more plausible than the extremes. Combining these possibilities into a single estimate gave a weighted average of 2.85 rides per user per month. With an estimated 609,930 Waymo rides in September and 2.85 rides per user per month, I estimated that roughly 214,000 San Franciscans used a Waymo that month.
I used Grow SF’s July 2025 survey (Results; archived on Dec 3, 2025)—the 30% of the 539 respondents in the sample who were classified as “supporters” of autonomous vehicles and cited “safer than human driving” as the reason for support—as a benchmark for a similar estimation. I assumed the 539 respondents were not perfectly representative of the estimated 214,000 San Franciscans who used a Waymo (by extension of Waymo being the only fully-deployed carrier in the city at the time, who used a self-driving car), and that anywhere between 20% and 50% of San Francisco Waymo users could have cited “safer than human driving” as their primary reason. I placed more weight on higher percentages (30–40%) than on the low end (20%), on the assumption that San Francisco Waymo (self-driving car) users are more likely than the general poll sample to cite safety as their main motivation. Specifically, I assigned 10% confidence to 20%, 35% to 30%, 35% to 40%, and 15% to 50%. This yielded a weighted average of 34% of San Francisco Waymo users who would have chosen Waymo (a self-driving car) primarily because it was safer than human driving. Applying this share to the estimated 214,000 users implied that roughly 72,760 San Franciscans chose Waymo (a self-driving car) primarily for safety.
GDP per capita numbers
The GDP per capita of San Francisco was calculated using the Federal Reserve Bank of St. Louis’s data: in 2023, the city’s GDP was reported at $263.1B ($263,108,633,000); the city’s population was 827.5K (827,526) residents (GDP; Resident Population; archived on Nov 14, 2025).
The GDP per capita of Singapore in 2023 was reported at $85,412 by the World Bank. (GDP per capita; archived on Nov 17, 2025). To validate the consistency of GDP per capita measurements across San Francisco (the St. Louis Fed data) and Singapore (the World Bank data), the GDP per capita of Singapore was also calculated by dividing the city-state’s GDP in 2023 by its total population that year. In 2023, the city-state’s GDP was reported at $505.4B ($505,439,514,078); the city-state’s population was 5.9M (5,917,648) residents (GDP; Population; archived on Nov 14, 2025 and Nov 16, 2025, respectively). Divided, the numbers produced $85,412, confirming the consistency of measurements.
The GDP per capita of Mostar was estimated as the midpoint between Bosnia and Herzegovina’s GDP per capita and Sarajevo’s GDP per capita. The underlying assumption was that Mostar’s GDP in 2023 had been lesser than that of Sarajevo, the country’s capital, but higher than the country’s average. First, the country’s GDP per capita in 2023 was reported at $8,663 by the World Bank (GDP per capita; archived on Dec 1, 2025). Second, Sarajevo’s GDP per capita in 2023 was calculated to be $17,828, derived by applying a conservative 2% growth rate over three years on the base of $16,800 GDP per capita, which was estimated by the Harvard Growth Lab (Metroverse data; archived on Nov 14, 2025), presumably for 2020 when the source also reported the city’s population. It’s worth noting, however, that the World Bank reported higher GDP per capita growth rates in the country for this period (GDP per capita growth rate; archived on Dec 1, 2025); these were not used because the base ($16,800) was already an estimation, hence the need for conservative growth rates. Finally, the GDP per capita of Mostar was calculated as the average between these two numbers—$8,663 and $17,828—at $13,245.
The GDP per capita of Mumbai was estimated by the Harvard Growth Lab (Metroverse data; archived on Nov 14, 2025) at $7,700. For context, the World Bank reported India’s GDP per capita in 2013 at $2,530 (GDP per capita; archived on Nov 17, 2025) and a 2024 report by the Economic Advisory Council to the PM reported the GDP of Maharashtra (the state in which Mumbai is) as percentage of India’s national GDP in 2023 at 13.3% (Report, archived on Nov 17, 2025). While not direct references to the $7,700 estimate from the Harvard Growth Lab, they provide some validity to the number, namely that it is higher than the country’s GDP per capita.
I don’t remember the first time airplane turbulence freaked me out, but I remember the day it turned into a nightmare, when it was no longer an unlikely event but a monster of tangible power and wickedness. The date was Oct 14, 2019, and I was on jetBlue flight B6 1715, flying from New York to San Francisco. The plane took off at 21:44 UTC-05:00 amid a heavy storm, and shortly after the city became a glistening dot in the darkness beneath us, violent blows began to thrust the plane.
I had never felt such a force from the skies. Initially, a few people—including me—were on high alert, but most passengers appeared unworried, some even amused by the rollercoaster ride. Suddenly, seemingly in a split second, our Airbus 321 aircraft was getting thrown in every direction possible. Up and down, left and right, clockwise and counterclockwise. By the tenth minute of these unrelenting blows, we were all thinking the same: we might die on this flight.
Some people prayed, some screamed, others vomited. One passenger, two rows in front of mine, spiraled into a panic attack and yelled in terror, asking what was happening, when it was going to stop. I clenched my fists, my face winced with each blow, my stomach twisted with each free fall. Like in those paralyzing dreams, I was not even able to shout.
I sent a text message to my mom on WhatsApp. I wrote that we were in very strong turbulence, that I had no idea what was going to happen, but that I loved her. I told her to tell my dad and brother the same. One checkmark appeared on the right side of my message, and then the wifi went out.
After about thirty minutes—an objectively short timeline that nonetheless felt like eternity—the turbulence subsided. Things went back to normal, and a few hours later, we landed safely in San Francisco. While we were taxiing on the tarmac, the second checkmark in WhatsApp appeared, then followed a reply from my mom, on whom the apocalyptic tone of my message was clearly lost, with a casual “honey, I was asleep. everything ok?”. There was never any explanation of what happened on that flight.
Turbulence has since become something I think about every time I have to get on an airplane. This obsessive thinking has manifested in a few ways. For instance, I always try to sit as close as possible to the aircraft’s center of gravity. Another example is that my seatbelt is always, always on, fastened almost hermetically around my waist. And if you were to judge me by the content of my carry-on backpack, you would probably think I was a survivalist.
The reactive approach has helped a bit, especially with turbulence seemingly getting worse over time once I started paying attention to it, but it’s only within the last year that I realized my self-structured exposure therapy (or at least, my self-awareness that I can’t avoid flying) would never be enough to calm my worries. I had to acquire some form of foresight, I needed to understand turbulence.
My friend, a fellow turbulence-fearer, suggested I start using turbli.com, a website that provides turbulence forecasts for any upcoming flight. The tool is neat—each trip is standardized by its duration on the x-axis, which gives a sense when in the trip one can expect turbulence, while the y-axis tells how strong the turbulence will feel when it happens. So, for instance, ahead of my flight from Taipei to San Francisco on March 9, 2024, the forecast on turbli.com looked like this (my interpretation of the chart from the website):
The app helps interpret the chart and tells how strong the turbulence might feel, but I still had no idea what any of this stuff actually meant. What was EDR, the Eddy Dissipation Rate? How did one know why some values were marked as light turbulence and others as strong? And then, obviously, what the hell even is turbulence? If I could study air turbulence, I thought—maybe, just maybe—I could conquer my newly acquired fear, by knowing the physics and math of this phenomenon, and I could approach my next flight through an investigative lens. Through a lens that would give me a sense of agency.
I became engrossed in this intricate world for months, trying to learn as much as possible before my flight from Taipei to San Francisco in early 2024, starting from turbulence as a physical concept and working my way up to the science behind the charts on turbli.com. What follows therefore is my best attempt to decode the nebulous aviation phenomenon that weighs on many travelers and, for reasons that I would learn only at the end of this endeavor, appears nebulous only because our own hubris made us believe we were entitled to any certainty in the first place.
Turbulence, as a concept, comes from fluid dynamics, a branch of physics that studies the flow of fluids. It might seem counterintuitive to be thinking of fluids in the context of airplane turbulence because air itself is a gas. Gas, however, is also a fluid because—like the other fluids that we typically think of when we say fluids, namely liquids—it can flow and take the shape of its container.
And just like most substances around us, air is made up of various molecules—primarily oxygen and nitrogen—with a few particles making guest appearances, such as water vapor and carbon dioxide. We can’t see air with the naked eye because there is a vast separation between its molecules, which leads us to think of air as nothing, perhaps an empty space, perhaps something that needs to be filled.
And yet, air is very much something. It is a fluid and it flows in all directions, and as a fluid, it can be deformed. When you wave your hand in the air, you deform it, causing the various molecules in air to disperse and change the direction of their flow. Like every fluid, air has an inherent resistance to this deformation, a fluid’s property known in physics as viscosity, better understood as the fluid’s “stickiness.”
As you have probably guessed by now, air is not sticky. The vast separation of molecules in the air means that air doesn’t have a lot of inherent resistance to deformation. When you wave your hand, you don’t feel anything preventing you from doing that, which is a sign of the air’s low viscosity.
The flip side of a fluid with low viscosity is that, once deformed by an external force, it doesn’t have enough “strength” to bounce back smoothly from this deformation and the fluid’s flow therefore becomes turbulent. Conversely, if a fluid has high viscosity—honey is a great example of such a fluid—it will be very resistant to deformation. Just imagine being inside a jar of honey and trying to wave your hand through the thick plasma. It would be near impossible.
What does all this have to do with planes? Let me try a visual, theoretical example. Imagine the air, a smooth jet stream, flowing west to east, without any interruptions on its way. Just the plain skies above and the plain fields below. This is what you see as the horizontal red lines in the image, on the left.
Physicists would call this laminar, sheet-like flow because the layers of the fluid are moving smoothly past each other. Then, a plot twist: the air’s flow encounters a nature’s massive obstacle on the way, like the Rocky Mountains in the United States or the Alps in Europe. This encounter will deform the air, and because the air is not viscous, it won’t be able to easily dampen the effect of this deformation. Enter turbulent flow, which are the red swirls seen around the mountain tops on the right.
The swirls are called eddies and they are the central characters in the agonizing story of turbulence. All sorts of causes—obstacles (like the mountain shown in the image), wind shear, temperature differences, and so on—create these eddies by applying stress to the fluid. In a way, eddies can be thought of as the unfortunate aftermath, an incurred cost perhaps, of the fluid’s low viscosity. When stress is applied to a low-viscosity fluid, that energy is injected into the fluid’s flow and, because the low-viscosity fluid is not great at resisting this process, its only way to respond to this chaos is to have swirling motions, eddies, to dissipate the energy. Hence the Eddy Dissipation Rate (EDR).
Planes get affected by eddies because an airplane is a three-dimensional object moving through a medium (air), and that means one cannot talk about flying in the skies without talking about the axes of an aircraft1, the imaginary lines that pass through the aircraft’s center of gravity at 90° angles of one another.
It’s normal for any three-dimensional object to move around these axes, in movements called yawing, pitching, and rolling. Pilots maneuver aircraft across the three axes to ascend, fly, and descend, which is all swell, but the air’s turbulent flow, with its swirling eddies, can also move the aircraft along these axes.2 So, to feel turbulence as a passenger on a plane is to feel the (sometimes violent) movements of the aircraft along these imaginary lines, caused by the air’s eddies.
Hopefully apparent by now is the inherent pecking order between air turbulence and an airplane caught in air turbulence. Turbulence happens, and it happens regardless of whether the airplane is flying through the region of air that has become turbulent, so it is not surprising that the formula used to calculate the Eddy Dissipation Rate—the turbulence metric supreme—has nothing to do with the aircraft itself.
In a 2012 research paper3, NASA researchers have outlined the following formula for calculating the Eddy Dissipation Rate (ϵ):
where e is the turbulence kinetic energy (TKE) and Le the integral length scale of turbulence; the former measures fluctuations of the (turbulent) three-dimensional velocity components from an average velocity, the latter measures the average size of the most energetic eddies within the turbulent flow.4
Rather unintuitive at first glance. To make it more palatable, it’s crucial to understand what this turbulence kinetic energy in the numerator really is. In classical mechanics, kinetic energy is usually expressed as the product of mass and velocity squared, denoting the importance of the singular object that possesses this energy due to its motion. In fluid mechanics, however, one studies the continuum of fluid particles, so to have valid comparisons and calculations of the phenomena within the fluid, it becomes necessary to standardize the kinetic energy by mass. In other words, to divide the unit of energy (Joule) by the unit of mass (kilogram).
Notice the lack of dependency on mass in the terminal expression once the units of mass are canceled out. In a way, the turbulence kinetic energy (e) is like an average, a statistical summary. But, a summary of whatexactly?
A summary of its raw data points, which are the three-dimensional velocity fluctuations mentioned earlier. Turbulence is, for all intents and purposes, chaos, which means there are numerous particles with localized velocity fluctuations from the average velocity of the fluid, and it’s these fluctuations that carry the kinetic energy. Expressing each fluctuation on its own is not helpful, so one takes an average of fluctuations in each of the components, and then also squares them to ensure bidirectional motions (which could cancel each other out) do not obscure the average magnitude of those fluctuations.
Once we account for the unit of the fluctuating velocity components, which is the same as the unit of velocity itself because we are simply measuring the difference of two velocity values, we get the same unit of turbulence kinetic energy.
Put simply, the turbulence kinetic energy (e) in the numerator is about the kinetic energy contained in the velocity fluctuations of the turbulent flow, which themselves are part of the turbulent swirls—the eddies. And, what about the denominator? The Le?
Here, it’s perhaps easier to think about the turbulence kinetic energy again. In the same way the kinetic energy of a singular object is not very helpful to the field of fluid dynamics, which is why we standardize it by mass to extract a new measure, that same turbulence kinetic energy on its own is not very useful to a pilot who has to fly through an apparently turbulent region of air. That’s because the turbulence kinetic energy doesn’t say anything about how that energy is distributed spatially. It’s not the same if the same amount of energy (standardized for mass) is spread over a tiny region versus a large one.
Intuitively, this is why the turbulence kinetic energy gets divided by the integral length scale, the average size (length) of the most energetic eddies in the Eddy Dissipation Rate formula. Revisiting the formula,
it’s obvious, albeit surprising maybe, that the Eddy Dissipation Rate and the integral length scale of turbulence are inversely proportional. Holding the numerator (the turbulence kinetic energy) constant, the smaller the integral length scale of turbulence, the larger the Eddy Dissipation Rate. Conversely, the larger the integral length scale of turbulence, the smaller the Eddy Dissipation Rate. So, if the most energetic eddies are really large, the Eddy Dissipation Rate will be really small. Uh, what?
Turns out, despite all the chaos inherent in a turbulent flow, the system—the fluid, and all the dynamic processes in it—works to keep things in an equilibrium. It’s the core principle of thermodynamics, that nature ultimately seeks to reduce instability. Larger eddies, by the nature of their size, allow “more room” for gradients of velocity and pressure across the fluid, which means the magnitude of forces acting at the interfaces between eddies and the surrounding (laminar) layers of fluid are smaller. In other words, smoother dissipation of energy, smoother transition. Smaller eddies, on the other hand, are the opposite of this: quick dissipation of energy, sharp gradients at the interfaces, more instability. It’s worth noting again: the energy in a turbulent flow does not come from the size of the eddies, but from the magnitude of the fluctuations, the magnitude of the gradients.
Knowing the unit of the integral length scale of turbulence, which is the unit of distance (meters),
we at last can derive the unit of the metric on the y-axis of turbli.com charts, the unit of the Eddy Dissipation Rate (EDR):
By doing this derivation, I realized that the EDR does not measure the probability of turbulence happening. The unit of meter squared per second cubed tells the pilots how turbulent—how chaotic—the atmosphere is, but, as is now evident from the derivations, that in itself is not a probability of turbulence happening, and, perhaps more interestingly, without any reference to the aircraft’s mass, it is not a direct indication of how the aircraft will experience the turbulence if it does happen. The latter confused me. I knew from my own experience as a passenger that bigger planes handled turbulence better.
This was the final piece of the puzzle. While the EDR is indeed not dependent on the aircraft’s mass, aircraft’s mass affects its own weight, because the force exerted on the aircraft by gravity is directly proportional to the aircraft’s mass.
The aircraft’s weight, in turn, is a critical component of the stall speed5, an aviation measure that signifies the minimum speed at which the aircraft must fly to stay aloft, and this part is a bit more intuitive—heavier aircraft has to fly faster than a lighter one to stay in the air. A higher stall speed ultimately means higher turbulence penetration speed, which is the greatest safe speed at which the aircraft can operate in moderately rough air6, and above which structural damage might occur in choppy skies.
Notice the implication here (albeit a bit simplified because many other factors are at play as well): the same turbulent atmosphere, in the same spot, at the same altitude, will appear “weaker” to a heavier aircraft because it will need more “strength” to counteract the plane’s high momentum, caused by its large mass and velocity.
turbli.com says that the EDR values (once multiplied by 100 for easier interpretation) translate to the following categorical turbulence classifications: light (0 — 20 m2/s3), moderate (20 — 40 m2/s3), severe (40—80 m2/s3), extreme (80—100 m2/s3). This means the app does not adjust turbulence classification according to the aircraft weight.
Other sources7 point out that these particular ranges are appropriate for medium-sized aircraft, like Boeing 737 and Airbus 320, whose maximum takeoff mass is between 15,000 lbs and 300,000 lbs. But for heavier aircraft, like Boeing 777 and Airbus 330, different ranges apply: light (0 — 24 m2/s3), moderate (24 — 54 m2/s3), severe (54 – 96 m2/s3), extreme (96 — 100 m2/s3). This would be of interest to me on my upcoming trip.
When the time came for my flight from Taipei to San Francisco on March 9, 2024, I opened turbli.com the day before, and saw the following chart:
Using turbli.com’s classification, without taking into account the model and the weight of the aircraft (Boeing 777-200ER), I overlaid the chart with the following ranges:
It seemed the flight was going to be slightly bumpy at the beginning, moderately-to-very and consistently bumpy in the middle, starting 5.5 hours into the trip and lasting for about an hour, and then very bumpy, though briefly, at landing. But I was curious how the weight-adjusted version of the chart would look if I took into account the maximum takeoff mass of Boeing 777-200 ER, classified as a heavy aircraft at approximately 600,000 lbs of maximum takeoff mass.
Adjusted for weight, the EDR chart for this flight would have the following overlay:
In this version, the trip should never be in the strong (severe) turbulence category, and even the biggest peak in the middle should feel mostly like light-to-moderate turbulence. The bump at landing should still feel notable, though not as severe as it would in a medium-weight aircraft.
On March 9, 2024, a rainy and foggy Saturday in Taiwan, UA852 took off at 13:19 UTC+08:00 from Taipei to San Francisco, and I, sitting in 35C, diligently documented how each hour of the flight felt, knowing that things would likely get scary five hours into the trip.
Flight Log: UA852 From: Taipei, Taiwan To: San Francisco, CA, USA Original Duration: 11h 14min Actual Duration: 10h 25m Departed at: 13:19 (UTC+08:00) / 21:19 (UTC-08:00) [9 min late] Arrived at: 23:45 (UTC+08:00) / 07:45 (UTC-08:00) [40 min early]
Hour 1: Slightly bumpy takeoff.
Hours 2—5: Smooth cruising.
Hour 5—6: It’s showtime.
02:30 UTC-08:00: Cabin lights on; flight attendants very quickly distribute snacks, clearly in a rush.
02:38 UTC-08:00: Pilot announces turbulence; jump seats secured.
05:12 UTC-08:00: Brief and light bumpiness; seatbelt sign back on.
Hour 9:
06:02 UTC-08:00: Some shakiness, but not like the big thump earlier.
Hour 10:
Smooth final hour.
Hour 10.5:
07:45 UTC-08:00: Strong but short burst of turbulence upon landing in SF.
The forecast was mostly accurate; everything just happened a tad earlier than expected, probably because the plane took a slightly different route once we got delayed or because the jet streams gave us an extra speed boost. It’s hard to say, however, which version of the categorical turbulence classification was better suited for my qualitative evaluation of the experience. The weight-adjusted chart, in which only the one-hour timeframe in the middle is classified as moderate turbulence seems more appropriate, though I probably would have answered differently in the moment.
I wish I could say this scholarly ordeal made things easier when the first frightening thump struck at 03:15 UTC-08:00. Sure, I felt somewhat comforted by the chart, knowing the turbulent episode wouldn’t last too long, but those five minutes still felt like an agonizing eternity. My stomach was still up in knots, my right hand still pressed tightly against the seat in front of me. I was still afraid.
When the turbulence settled and the flight attendants turned off the lights, I kept thinking, at the edge of sleep, about this predicament of mine. It occurred to me that I was not afraid of death from extreme turbulence, that’s not what made me feel so uneasy. Make no mistake, I never felt indifferent about it. I am always deeply vested—I might say even passionate—in making it to my destination in one piece and continuing on with the minutiae of my daily life on planet Earth. But, you know, if it’s meant to be my time, then it’s meant to be my time.
The discomfort, that dreadful feeling of doom in my gut that creeps in each time turbulence strikes, comes from the liminal time period between the onset of turbulence and its end, which seems to continuously slip away, trapping me in a quantum superposition—like Schrödinger’s cat—both alive and dead, awaiting the outcome dictated by forces beyond my control.
It comes, perhaps, from the the sobering reality of my—and our—poetically fragile existence, a stark reminder that, despite the technological advancements we make, despite the reassuring air travel statistics we cite, despite the scientific frameworks we impose on the world around us, when we confront nature’s fury and greatness, we also confront our defenselessness and insignificance.
Never clearer is this universal truth than at the precipice of moderate to strong turbulence, like in that split second on my jetBlue flight in 2019, when even the most phlegmatic of us all, previously unfazed by the atmosphere’s violence, shriek in terror. It’s at this moment that we confront our lack of agency in the grand scheme of things.
As nihilistic as this all sounds, that day, in that moment, on my way from Taipei to San Francisco, in the suspended state of sleep, between lucidity and delirium, it somehow made the monster feel less scary, as if I realized that it had always been there—I just never acknowledged it.
Factual references
U.S. Department of Transportation, Federal Aviation Administration, Airplane Flying Handbook(2021), FAA-H-8083-3c, Glossary G-2. ↩︎
National Weather Service, ZHU Training Page — Turbulence.↩︎
U.S. Department of Transportation, Federal Aviation Administration, Turner, T. Flying Lessons for May 6, 2010, (copyright of Mastery Flight Training), pages 1-2. ↩︎
Aviation Weather Training (AvWxTraining), EZWxBrief Pilots Guide. (No color version, updated 5/28/2024), Version 2.0.0, pages 47-48. ↩︎
Part 1 of this post outlined the three steps—and the three corresponding decision trees—that every analytics professional in the B2B SaaS 1 space should use when evaluating a project proposed by other business stakeholders. In many cases, those simple but effective tools reveal that the stakeholder’s proposed project is not worth doing. But, what happens when the three decision trees all signal the project mightbe worth doing, as shown below?
Step 1: Is it obvious what the right decision is, even without this analysis / model / investigation?
Answer 1: No. → Decision 1: The project might be worth doing.
Step 2: Will this analysis / model / investigation actually be used to make a decision?
Answer 2: Yes. → Decision 2: The project might be worth doing.
Step 3: If instead of doing this analysis / model / investigation, you did a “back-of-the-envelope” calculation, would you get ~80% of the insight you need?
Answer 3: No. → Decision 3: The project might be worth doing.
Then, it’s time for the fourth step: calculating the benefit-cost ratio of the proposed project. It’s one way to do what analytics (and product) professionals call opportunity sizing, though it might be confusing if we use the term opportunity sizing because, in this context, the decision is guided by the (financial) bottom line of resourcing full-time employees. That is, if staffed on the proposed project, will the financial cost of the analytics professionals’ salaries be offset by future financial gains for the business from completing this project? It’s really about the financial viability of the project from a resourcing perspective.
Step 4: Estimate the benefit-cost ratio of the proposed project and make a final call
4a: The benefit-cost ratio
The benefit-cost ratio can be used to estimate the cost-effectiveness of a proposed business project. It’s useful only if both the value of the benefit and the value of the cost are expressed in the same unit, which, in the context of evaluating business projects, should be a financial unit, like USD ($).
Analytics professionals who are evaluating proposed business projects can use these definitions of benefit and cost:
Benefit ($): amount of labor cost decrease or net new revenue this project would help generate (summed if both available)
Cost ($): amount of money the business will spend on the analytics professionals to do this project
A project is considered cost-effective (in other words, worth doing) if it’s greater than 1 because the benefit is higher than the cost. That’s the textbook definition of cost-effectiveness, but in many real-life business scenarios, this threshold will need to be higher, like 5 or 10. This is especially true if the business is cash-strapped and needs to invest in only the most valuable projects.
Step 4 takeaway: The benefit-cost ratio can be used to quantitatively assess if a proposed project is worth doing. It’s a dimensionless ratio calculated by dividing the estimated benefit of the project by the estimated cost of the project. The textbook definition says the projects worth doing always have a benefit-cost ratio greater than 1, but, in reality, the threshold is usually higher than that.
4b: How to estimate the benefit ($)
Calculating the true benefit ($) of a project in B2B SaaS space can be complicated. Do you consider impact on the overall sales motion? Or just new-business deals? Renewals? Expansions? Upsells? How do you know much credit to give to the analytics professionals; after all, they are not the ones facing the customers?
For this particular purpose, it’s okay to estimate and make assumptions, because the analytics professional has to make a quick decision on whether to say yes to a project. Let’s use the definition provided earlier and assume that the benefit ($) of a proposed business project can be estimated as the sum of estimated labor cost decrease and estimated net new revenue.
To estimate the labor decrease of the project, we need to provide best estimates for the following four values:
approximate number of business stakeholders trying to generate these insights manually (the assumption here is that the business stakeholder wants help from the Analytics team because the business stakeholder can’t easily generate these insights independently)
approximate number of hours per week a single business stakeholder spends doing this manually
(composite) hourly salary of the business stakeholder(s)
number of working weeks in a year that the business stakeholder(s) spend(s) doing this task
Once we have the best estimates, we multiply the four values as shown in the image below:
The idea behind this calculation is that it represents how many business stakeholder hours (and therefore, how much of the business stakeholder salary) would be saved and invested elsewhere if the analytics professional did this project. For instance, if a business stakeholder is spending six hours each week visualizing data in Google Sheets, an analyst could build a dashboard in a BI tool like Looker or Tableau, which would free up the business stakeholder’s time because the visualizations are now automated in the dashboard.
The second part of the benefit equation is the net new revenue this proposed project would bring. Again, we could make this a complex equation by accounting for all nuances (do we think in terms of customer accounts or individual deals? how do you account for multi-year deals with renewals? do you account separately for new-business deals versus expansion deals? do you handle each type of expansions separately?), but for quick decision-making, we can provide best estimates for these three values:
estimated number of net new accounts this project would help bring in through sales
estimated number of accounts this project would help prevent from churning (objectively, these accounts are not net new, but let’s ignore that technicality to make the calculation easier)
average annual (customer) contract value ($)
Once we have the best estimates, we multiply the ACV ($) with both the estimated number of net new accounts and the estimated number of accounts saved from churning as show in the image below:
Notice that I am glossing over some of the mathematical imprecision, like using contract-level average annual values and multiplying them with the number of customer accounts, so it’s unclear which level the values are averaged across (over all contract values? over accounts’ average contract values?). Because it’s a quick calculation, let’s assume they are the same.
The idea behind this calculation is that it considers whether the proposed project could help generate revenue. It’s best to be conservative in this case because it is highly unlikely that the analytics professional’s project on its own would generate revenue. Instead, the project would make the business stakeholder (like the customer success manager or the sales rep) more effective at their job, which would presumably make it easier for them to impact the go-to-market motion. Two ways to make it conservative: (i) either use a low number of accounts, or, more intuitively, (ii) multiply the final value by a percentage (%) that represents how much credit the analytics professional can take in generating net new revenue.
Step 4b takeaway: To estimate the benefit of the proposed project, the analytics professional needs to estimate the labor cost decrease and the net new revenue that the proposed project would generate. Some projects will generate one or the other, and some projects will generate both. Automation of manual work (for instance, making a dashboard or a “click and run” Python notebook script) is what usually drives labor cost decrease. New tools and capabilities (like predictive models that inform go-to-market or product strategy in a data-driven way) are usually what drive net new revenue.
4c: How to estimate the cost ($)
Calculating the cost ($) of the proposed project is a bit easier.
We need to provide best estimates for the following two values:
(composite) hourly salary of the analytics professional(s)
estimated number of hours that the analytics professional(s) would spend on this project
If the analytics professional evaluating the project is the only full-time employee who would work on this proposed project, then finding the hourly salary is straightforward (divide annual salary by the number of working weeks in a year and by the number of working hours in a working week). If there are other full-time analytics employees involved, then it’s best to use a composite hourly salary of all the analytics professionals, and account for all of their hours. The composite hourly salary can be estimated by finding the median annual salary across these roles using a tool like Glassdoor or levels.fyi.
Step 4c takeaway: Calculating the estimated cost of the project is usually easier than calculating the benefit. The analytics professional needs to know the projected number of hours the Analytics team would spend on the project and the (composite) hourly salary of the teammates working on the project. Hours spent on project maintenance, stakeholder enablement, as well as any post-mortem work should also be included in the projections.
4d: The units of benefit and cost
This step is purely informative, in case you are like me and like to map units to each metric when running calculations.
For metrics behind the benefit calculation, we have:
approximate number of business stakeholders trying to generate these insights manually [person]
approximate number of hours per week a single business stakeholder spends doing this manually [hr / week / person]
(composite) hourly salary of the business stakeholder(s) [USD ($) / hr]
number of working weeks in a year that the business stakeholder(s) spend(s) doing this task [week]
estimated number of net new accounts this project would help bring in through sales [account]
estimated number of accounts this project would help prevent from churning [account]
average annual (customer) contract value ($) [USD ($) / account]
For metrics behind the cost calculation, we have:
(composite) hourly salary of the analytics professional(s) [USD ($) / hr]
estimated number of hours that the analytics professional(s) would spend on this project [hr]
The image below shows that these units cancel each other out and that we end up with a dimensionless benefit-cost ratio.
Let’s demonstrate why benefit-cost ratio works. We can do that by stress-testing it with two extreme scenarios, for which it would not even be necessary to calculate the BCR: (i) a project that is obviously not worth doing and (ii) a project that is obviously worth doing.
4e: Stress-testing the benefit-cost ratio: Extreme scenarios
In the first scenario, a product manager asks if someone from the Analytics team could build data warehouse pipelines (net new data tables using ETL tools) and then a net new business-intelligence dashboard that would help the product manager track test data in preproduction environment produced by engineers as part of a larger year-long initiative to revamp how clickstream events are instrumented. According to the product manager’s roadmap, the product manager will spend close to 2 weeks every quarter, and about 2 to 3 hours per week, exploring and testing the preproduction data, and it would be great if the product manager could access the preproduction data easily in a business intelligence tool.
Why is this obviously notworth doing? Well, it is understandably an inconvenience for the product manager to have to query the test data in the preproduction environment each quarter, having to potentially rewrite code or adjust it, download the data, and do visualizations in a tool like Google Sheets. At the same time, it is preproduction test data—it’s used for just this project, just to quality-check the data for potential issues, and it will be used for at most two weeks each quarter, which means for at most eight weeks in the entire year. In addition to all that, preproduction data tends to be patchy and sparse. To invest in production-like data architecture for this proposed project would clearly be a poor decision.
The benefit-cost ratio is not even necessary in this scenario, but it can be used to illustrate why it would not be wise to work on this project. Let’s use the following best estimates (for hourly salary, we will use a simple placeholder of $50/hr for all full-time employees and for average annual contract value, we will also use a placeholder of $500K/account]—
Values needed for the numerator (benefit):
approximate number of business stakeholders trying to generate these insights manually [1 person]
approximate number of hours per week a single business stakeholders spends doing this manually [3 hrs / week / person]
(composite) hourly salary of the business stakeholder(s) [$50 / hr]
number of working weeks in a year the business stakeholder spends trying to generate these insights manually [8 weeks]
estimated number of net new accounts this project would help bring in through sales [0 accounts]
estimated number of accounts this project would help prevent from churning [0 accounts]
average annual (customer) contract value ($) [$500K / account]
Values needed for the denominator (cost):
(composite) hourly salary of the analytics professional(s) working on this project [$50 / hr]
estimated number of hours that the analytics professional(s) would spend on this project [~1 working week = 40 hours]
With a benefit-cost ratio of 0.6, the project would produce net loss for the business, and is therefore not worth taking on.
For the second scenario, let’s imagine the Sales team asks if the Analytics team can build a predictive model that scores new-business accounts in the sales pipeline for their likelihood of closing. The team wants to use the data to allocate full-time employee resources strategically; right now, seemingly promising deals fall through at the last moment, and the team is pulling data from business-intelligence tools every week to no avail—the descriptive data is clearly not helping them. According to the team, about 30 sales reps are trying to make sense of the descriptive data each week, and each rep spends about 8 hours per week collecting the data from business intelligence tools, stitching different metrics, and trying to make sense of what the data is saying. Put differently, it’s crisis mode for the Sales team.
Why is this obviously worth doing? First, it’s every analytics professional’s dream opportunity: to build a predictive model that will be used continuously by the business to make strategic decisions and that will provide unique intelligence to the business that otherwise could not be obtained. Just imagine if the predictive model was built and implemented successfully: so many hours of unnecessary work would be saved for sales reps and appropriate resourcing around the pipeline accounts would ensure more deals are closed. It’s most certainly a hefty investment for the Analytics team because it would require exploratory analysis, model building, testing, and then enablement. But, that’s exactly what the team is uniquely positioned to do.
The benefit-cost ratio is also not necessary in this scenario, but it can be used to illustrate why this project should definitely be prioritized by the Analytics team. Once again, let’s use the following best estimates (for hourly salary, we will use a simple placeholder of $50/hr for all full-time employees and for average annual contract value, we will also use a placeholder of $500K/account]—
Values needed for the numerator (benefit):
approximate number of business stakeholders trying to generate these insights manually [30 persons]
approximate number of hours per week a single business stakeholders spends doing this manually [8 hrs / week / person]
(composite) hourly salary of the business stakeholder(s) [$50 / hr]
number of working weeks in a year the business stakeholder spends trying to generate these insights manually [~A full working year = 48 weeks]
estimated number of net new accounts this project would help bring in through sales [~5 accounts]
estimated number of accounts this project would help prevent from churning [0 accounts]
average annual (customer) contract value ($) [$500K / account]
Values needed for the denominator (cost):
(composite) hourly salary of the analytics professional(s) working on this project [$50 / hr]
estimated number of hours that the analytics professional(s) would spend on this project [~A full quarter = 12 working weeks = 480 hours]
With a benefit-cost ratio of 128.2, which is hyperbolic because this is an extreme scenario, the project would produce notable net gain for the business, and is therefore worth taking on. Most projects will never have a benefit-cost ratio this high, but in this example, the ratio illustrates the magnitude of the proposed project’s benefit to the business.
Step 4e takeaway: If you are calculating a benefit-cost ratio, it’s very likely that it is not obvious if the project is worth doing. Very few projects will have BCR ratios as low or as high as the two illustrated extreme scenarios; instead, most BCRs will hover between 1 and 10, or potentially between 10 and 20. If the calculations are consistently producing high BCRs, it’s likely that the benefit estimates are too liberal, and this is most often a result of overestimating net new revenue.
5: The full “When to say no” decision tree
We now have a full decision tree for when to say no to a proposed analytics project in the B2B SaaS space (the thresholds for Step 4 will vary depending on the context of the business):
Step 1: Is it obvious what the right decision is, even without this analysis / model / investigation?
Answer 1: No. → Decision 1: The project might be worth doing.
Step 2: Will this analysis / model / investigation actually be used to make a decision?
Answer 2: Yes. → Decision 2: The project might be worth doing.
Step 3: If instead of doing this analysis / model / investigation, you did a “back-of-the-envelope” calculation, would you get ~80% of the insight you need?
Answer 3: No. → Decision 3: The project might be worth doing.
Step 4: Run a benefit-cost ratio calculation for the proposed project.
Answer 4: BCR less than 1.0 → Decision 4: The project is not worth doing.
Answer 4: BCR equal to or greater than 1.0 and less than 10.0 → Decision 4: The project is worth doing, but should not be a top-priority project.
Answer 4: BCR equal to or greater than 10.0 → Decision 4: The project is worth doing and should be a top-priority project.
At first, this might seem like a tedious process , but after a few trial runs, it starts to feel natural and, most importantly, it reduces guesswork.
Hyper-condensed acronym for “business-to-business (B2B) software-as-a-service (SaaS).” ↩︎
Since there is so much buzz around data these days in B2B SaaS companies1, especially with the latest insurgence of AI, many business professionals often turn to their analytics colleagues in search of insights that will help make the right move and do the right thing.
Sometimes, that’s exciting because the business really does need data to make the right choices. But sometimes, data can be just a scapegoat for many other unaddressed issues in the business: lack of alignment, lack of communication, lack of clear strategy, overcomplicating, and sometimes even laziness.
What no one ever told me was that a big part of being a strong analytics professional in these situations is not only knowing how to build the right data models and do the right analyses, but knowing when not to do any analytics whatsoever, and when to say no to data projects.
To be more precise, if you can evaluate early on whether a potential analytics project will trulymake an outsized impact on the business, saying no to low-value projects is actually much better for the company in the long run. It makes you a strategic thinker and a more valuable asset to the business.
That’s why I want to share my approach of evaluating potential analytics projects. It’s not perfect by any means but it has served me reasonably well in my day-to-day job.
It consists of four steps:
Step 1: Check whether you even need analytics to make a decision
Step 2: Determine if the business stakeholder(s) will actually use your work to make a decision
Step 3: Evaluate if you can get to (approximately) the right answer by doing a back-of-the-envelope calculation instead of full-scale analytics
Step 4: If the project passes the first three checks, then use the benefit-cost ratio to make a final call
The first three steps are judgment calls. The fourth step is the only one that involves math and that can be a bit time-consuming at first. Once you do it a few times, it becomes much quicker, especially if you make an automated version of it in Google Sheets or Excel. I describe the first three steps in this post, Part 1, and focus exclusively on the fourth step in Part 2.
Note: This approach is based on my personal work experience at high-growth, fast-moving B2B SaaS businesses, where my focus has usually been on driving strategic business decisions with data. My perception, also, is that analytics in these environments has a higher tolerance for estimates and assumptions compared to analytics in established, process-oriented companies. Keep that in mind as you review these steps because they might not be applicable to all analytics teams.
Step 1: Check whether you even need analytics to make a decision
As counterintuitive as this sounds, the first step when taking on a potential analytics project is to figure out if analytics is even needed to make a business decision. Data is useful when the right decision is not obvious, but it can be frustratingly wasteful when it’s used for every business decision.
This happens often because people think they need to see data for everything. In itself, that frame of thinking is a good thing, because it means people are knowingly trying to be more data-driven in their decision-making at work.
The thing is, you can sometimes make a very educated decision without any data, and doing so when appropriate is one of the most powerful skills in the arsenal of any data scientist or analyst. If you can help your business stakeholder recognize that the right decision is obvious and that no analytics is actually necessary, that’s far more valuable to the business than doing an impressive, highly rigorous analysis.
Here’s an example. Let’s say you are working as a data scientist for an innovative B2B SaaS company that sells all-in-one HR software (payroll, performance reviews, taxes, etc.) to other businesses. Every year, these businesses—your company’s customers—decide whether to renew their contracts and continue with your company’s services. The company has been operating for a while, so you have at least ten years of renewal and churn data at your disposal.
The software has a specific section dedicated to taxes, and the version that’s shown to US clients is phenomenal: slick, concise, easy to use, and even entertaining. US clients love the tax section and ever since it has been introduced four years ago, they have been highlighting it as one of the features that makes your company stand out among other HR tech businesses.
But the European clients don’t have that. Their tax section was done haphazardly many years ago when your company had only three clients in the EU, and it now looks like a sad relic of the past. It’s slow, confusing, and jarringly unpleasant compared to the other sections, which are great and equivalent to the ones shown to US clients.
You weren’t aware of this difference, but it has been brought to your attention after a very data-driven product manager joined the company. This product manager is evaluating which parts of the product need to be improved, and they come to you with this question:
“I am looking into evaluating which parts of the product need to be improved to elevate the customer experience. The European version of the tax section is super janky, but I would love to understand whether it has historically impacted financial outcomes, and if we can infer historical impact on customer churn. And, if so, what was the magnitude of that churn? I know that we collect customer satisfaction data through the NPS survey, so I wonder if we can use that to draw correlation or predictiveness? Having data-driven evidence would be super helpful to decide if my team should revamp the European version of the tax section.“
All sorts of questions come to mind when you see a request like this from your business stakeholder. What if we don’t have enough NPS responses? Do we even have enough European clients to make a legitimate analysis? How do I tell them that we can’t attribute churn to just the poor quality of the tax section? In a situation like this, the wisest approach is to pause and ask yourself if the right decision is obvious, even without any data.
I would argue the decision is obvious in this case. This is a typical scenario of overthinking and overcomplicating. The European version of the tax section has to be revamped. Whether its poor user interface was ever correlated to churn is irrelevant. It is a poor business practice if clients from one market have the good version of the product and clients from another market have the bad version of the product. It also introduces an implicit customer bias based on something the client can’t control. Not to mention that it can impact the company’s reputation if this discrepancy goes on forever.
The product manager does have to decide eventually how long this revamp should take and how to prioritize it compared to other projects, but they don’t need data to know that the revamp has to happen.
Step 1 takeaway: The answer to important business questions is sometimes simple and obvious. Knowing how to deconstruct your stakeholders’ requests and how to identify decisions that need no data is an underrated analytics skill.
Step 2: Determine if the business stakeholder(s) will actually use your work to make a decision
The more common scenario is that in which the decision is not obvious. It still doesn’t mean you have to get into the data immediately. You should also suss out if your business stakeholder is even going to use the insights you generate. By “use,” I don’t mean whether they will reference your work in a conversation with an executive, but whether your insights will actually drive decision-making in the business. How do you recognize this difference? By watching out for the greatest red flag in the world of analytics—the “it would be interesting to know” statement.
Here’s an example. We’ll use again the case of the B2B SaaS company that sells all-in-one HR software. Let’s say an account manager comes to you and says that the customers (in this case, the buyers of the HR software on the side of the client company) often ask whether usage personas can be identified among their employees by looking specifically at how their employees use the optional peer feedback tool provided by your company’s HR software.
This peer feedback tool is purchased as part of a (slightly) higher-priced bundle of your company’s product, and has generally been praised by customers for its understandable, AI-powered interpretation of qualitative text. That said, it is optional, so customers do sometimes wonder if they’re getting their bang for the buck.
This could certainly be a cool analysis, you think to yourself. Maybe those employees who use the optional peer feedback tool are also more likely to have better performance reviews from their managers? Maybe they are the employees who end up getting promoted more often? This would require thoughtful analysis setup and data exploration, but you definitely see promise in the account manager’s idea. In this situation, it’s very important to probe further, and to understand how this analysis would be used in practice.
“That’s definitely a great question,” you say. “There could be a bunch of interesting correlations and potential implications there. Is the idea that the people on the customer side, who run the HR departments, would then nudge those employees who don’t often use the peer review tool to use it more frequently? In case there is a positive correlation? What would happen if there is no correlation or if it’s even maybe negative?”
“Um,” the account manager sighs, “well, I don’t know that they would necessarily do anything. That really depends on their company’s policy and how the peer review tool is implemented and on the professional relationships at the company. But, it would be so interesting to know. I have so many customers asking this question. This insight could really elevate our narrative.”
Bam. That could have been several days of work for you: work that would have resulted in no identifiable progress for your company. When something like this happens, it’s okay to say no until the business stakeholders can identify clear motivation for this type of work and a robust course of action that would follow from the analysis.
Step 2 takeaway: Being curious is fantastic, especially if curiosity is a prominent quality in your company’s culture, but you need to remember that you are getting paid to help run a business. If you establish that your work will be used to simply generate an interesting insight for someone and will never be used for any decision-making, you will waste the company’s time and money if you say yes simply because the work sounds impactful.
Step 3: Evaluate if you can get to (approximately) the right answer by doing a back-of-the-envelope calculation instead of full-scale analytics
Okay, so let’s say the decision is not obvious and this potential project will definitely drive decision-making across the business. Well, that’s it. Sounds like you should go build that predictive model, kick off that in-depth analysis, or launch that full-blown data investigation, right?
Not necessarily. A strong analytics professional will always evaluate whether most of the needed insight can be accurately approximated with a back-of-the-envelope calculation before they commit to a full-scale analytics project. This particular step has been the hardest one for me to learn, because it goes against my perfectionistic tendencies and my selfish desire to always work on something new, challenging, and rigorous. But building this muscle—the intuition to recognize diminishing returns in potential projects—has been paramount in my analytics career.
Here’s an example. We’re still working at the B2B SaaS company that sells all-in-one HR software. The optional peer feedback tool is still the hot topic among customers, and this time, another account manager comes to you to ask for some data. They are elated because one of their accounts is so close to signing a multi-year, 5-year renewal deal. Unprecedented! But—the buyer on the customer account would love to know how this peer feedback tool thingy works longterm, in case they commit to the bundled product for the next five years.
For the buyer, the decision is clearly not obvious. Maybe people grow tired of the peer feedback tool after three years? Plus, any work you do here would clearly help the business; it might be the last insight needed to lock in this unprecedented deal. So, when the account manager comes to you and asks if you can do a statistical analysis that shows whether strong usage patterns of the tool in the first few quarters indicate likelihood of continued usage in the last six months of a deal, it might be tempting to roll up your sleeves and immediately get into the weeds.
Before you do that, first assess what you already know and have from a data perspective:
(a) You know that you don’t have any other five-year deals as means of comparison, if you were to perform a statistically sound analysis. Two years is the longest multi-year renewal your company has had so far. That said, it doesn’t mean you can’t extrapolate.
(b) From previous analyses you and other analytics teammates have done, you know that for the two-year deals, on average, 85% of the customer’s employees request peer feedback in this tool at least once in a 12-month period.
(c) Of those 85%, a typical customer employee requests peer feedback in this tool on average 1.5 times per month.
(d) You also know that requests are very frequent in the first and the last quarter of the year, and are less frequent over the summer months in the third quarter.
Then you also want to assess if there are sensible assumptions you can make about the length of this potential 5-year deal. Particularly:
(e) Five years is a long time. In the modern economy and especially in the tech industry, many employees will not stay at one place for 5 years. If someone joins the customer company at the beginning of this deal, it is likely they will leave the customer company and go elsewhere by the time this multi-year deal is over.
(f) Company leadership will change in those five years. Employees will switch teams and work with new peers in those five years. All to say, things will keep changing and will not stay the same in that 5-year period. Which means it’s likely they will want to keep hearing feedback from their (new) peers.
Knowing all this, you realize that you might already have a lot of the info you need to give a reasonable recommendation. So, you tell the account manager:
“Thanks so much for your request! I have noodled on this for a bit, and I think we can derive an answer that might be helpful to the customer. Here is how I would break it down:
(1) It is hard to have a statistically sound analysis around this because we have never had a 5-year renewal deal before. It’s also impractical to project correlations between early and later usage across such a lengthy deal because (as I’m sure the buyer will be also aware) so many things can change in five years! Employees will leave, they will change teams, they will get new managers, and hopefully many other good changes will happen.
(2) What we do know is that for the two-year deals we’ve sold, 85% of employees use the peer feedback tool at least once in 12 months, and of those 85%, an employee requests peer feedback on average 1.5 times a month.
(3) We expect that the 5-year deal will follow similar patterns, and that the customer will see lulls during the summer months each year, which is expected.
(4) The customer can also expect that not all of their employees will stick around for 5 years, so we can assume that we will go from 85% to maybe only 30—40% employees who will have used the tool at least once every year in the 5-year span.
(5) For that subset of users, we can assume that their average usage might taper off in the fourth and fifth year, to around 1 feedback cycle request every 2 months, but that will also be offset by more frequent usage from new employees who joined after the 5-year deal was signed.”
Tada! Yet again, you saved time and money for the business by not spending days on answering this question. This approach might feel hand-wavy, but think about the other scenario, in which you do a full-blown statistical analysis for the business. Would you arrive to a notably different conclusion? Likely not. Your analysis would probably show something similar, and even if you identified the more nuanced insight that employees who engage with the peer feedback tool frequently early on are also more likely to engage with it later on, you would still face the same conundrum—does it hold four to five years in? Probably not, for all the same reasons you already knew.
Step 3 takeaway: While it might feel unrigorous and uneasy, always ask yourself if you can extract most of the needed insight (“80% of what you need”) with a back-of-the-envelope-calculation (“20% of the work”). If the answer is yes, resist the temptation to do full-scale analytics. You will otherwise be delivering diminishing returns to the business. Save the time and energy for business problems that will truly require the precision and the rigor.
Those are the first three steps! Even if the project has passed all the three checks, we’re still at a “might be worth doing.” In Part 2, I cover the fourth step—the more rigorous step—that helps us make the final call on whether a “might be” should become a “definitely”.
Hyper-condensed acronym for “business-to-business (B2B) software-as-a-service (SaaS).” ↩︎
The first time I read Joan Didion’s “Slouching Towards Bethlehem” was in 2017, which was coincidentally also the year of the essay’s 50th anniversary. I remember feeling transfixed by it, partly because of Didion’s incisive writing and partly because of the striking characters that Didion met on the streets of San Francisco in the sixties. Characters that felt familiar, somehow almost too real.
I have reread the essay numerous times since then, and each time, the San Francisco of Didion’s world has felt even more similar to the San Francisco of my world. The longer I live in San Francisco, the more prophetic Didion’s essay becomes, and as I approach my seventh year of living in the city, the more convinced I am that the San Francisco I know today is an uncanny reincarnation of the same social fabric Didion witnessed in the sixties.
Joan Didion in SF’s Haight-Ashbury neighborhood, 1967. Photo courtesy of Ted Streshinsky.People gathering in SF’s Civic Center, 2022. San Francisco Love Tours van in SF’s Lower Haight neighborhood, 2022.
The characters Didion meets in the city are unmistakably Californian. Not in the sense that they grew up in California but in the sense that they came to California to find their true selves, only to later realize they didn’t even know what they were actually searching for.
There is Max, who tells Didion he “lives free of all the old middle-class Freudian hang-ups,” seeing his life as a triumph over “don’ts,” evidenced by his adolescent affairs with peyote, alcohol, mescaline, Methedrine, and his resistance to committed relationships. The same Max who later invites Didion to his place for a group trip on acid, but tells her that they have to wait for six to seven days because he and another friend, Tom, have been on STP for a while. The same Max who, as Didion later learns, is a trust fund baby and plans to travel to Africa and India to “live off the land” with his girl Sharon.
Sharon, a teenager who left her separated parents and moved to the Haight-Ashbury neighborhood, is keen to transcend the banality of everyday life by getting high, even making a door sign that said “DO NOT DISTURB, RING, KNOCK, OR IN ANY OTHER WAY DISTURB. LOVE.” so that the group could drop acid in peace. But she gets restless on the day of, waiting for Tom and his girl, Barbara, and is bored “just sitting around.” Max tells Didion that Sharon exhibits “pre-acid uptight jitters.”
There is also Don, who is on a macrobiotic diet. Then there is Jeff, who doesn’t pre-plan and lets “it all happen” and who really doesn’t like the fact his mother would ground him for not ironing his shirts for the week. He describes her as “just a genuine all-American bitch.” Another gem is a guy named Sandy, who is reading Laura Huxley’s You Are Not the Target when Didion first meets him and who sees meditation as a turn-on. Or a guy who goes by the name Deadeye, who “made a connection” to earn money and prevent getting evicted from his house, by getting acid from someone who had it and giving the acid to someone who wanted it.
It’s hard for me to read Didion’s essay and not marvel at the remarkable similarity, even fifty years later, between her protagonists and the people I regularly meet in San Francisco. Similar evangelists abound today in the city, be it the two guys I’ve known who once competitively compared their meditation routines, the friends who worship at the altar of Soylent, or the trust-fund acquaintances who microdose and read self-help books to find purpose amid the noise of the Bay Area rat race.
The industry I work in is predicated on people’s triumphs over “don’ts,” so much so that every veteran techie will proudly admit their startup idea was initially met with discouragement from their family, from their friends, and from their investors. It’s a badge of honor, in the tech industry, to resist society’s disbelief and to endlessly explore and play, not unlike how a headstrong kid defies its parents and does not yield to house rules.
That same resistance is highly valued today in San Francisco’s social contracts as well. A girl I recently met at a friend’s party moved from the South Bay to San Francisco in her early twenties, and found herself inducted into an unofficial polyamorous commune, which she described as a friend group at a “very thin line between a community and a cult.” She loved it because the people there helped her break out of her shell, but she did admit that monogamy was not less preferred—it was implicitly discouraged.
Haight-Ashbury, San Francisco in the sixties. Photo courtesy of Jim Marshall Photography.Castro neighborhood, San Francisco in 2021.
When Max, Sharon, and Tom drop acid in the essay, lounging together with Didion, there is a beautiful moment of fabricated, childlike togetherness. After all the “innumerable last-minute things” Tom has to do, Sharon’s pre-acid uptight jitters, Barbara’s indecisiveness over whether to smoke hash or drop acid, the group finally gets to enjoy the high, and there is no sound or conversation until four hours later, when Max simply says, “Wow.”
This hyper-optimized communal spontaneity that Didion’s new friends continuously chase strikes me as not so different from the togetherness that some of my friends and people I know in the city diligently engineer every year at Burning Man. One night in 2019, after telling a friend that I would never go to Burning Man because spending several days in the desert without all my skincare and fragrances sounds like hell, the same friend complained how I don’t seem to understand the connection and love she feels with other Burners for that one week in the desert.
I did want to ask her why that connection and love couldn’t exist outside the playa, but I figured I already knew the answer: when playing hard at Burning Man was over, she had to go back to working hard. In the words of one workaholic, energetic type-A tech executive who once told me before leaving for Burning Man, he was to be “off grid” for customers as he set out to playa to “lose a few brain cells.”
Don’t get me wrong, I don’t think I am above this. As Didion mentioned in the documentary Joan Didion: The Center Will Not Hold, “Killing a snake is the same as having a snake.” I don’t go to Burning Man and I don’t drink Soylent, but I know that I am also a consumer of the bumptious lifestyle that comes with the mainstream San Francisco scene. And I also know that my view of the city is myopically limited to the tech scene, which is where I have ended up by the mere nature of working in that industry.
What I do want to point out instead is that maybe the veneration and the scorn surrounding the social scene today in San Francisco are the same opposing forces that the city had seen before. Maybe these seemingly disparate communities—the hippies of the sixties and the techies of today—actually have more in common than one would think.
There is a good question presented in the essay, from a psychiatrist Didion met in San Francisco, that I found particularly indicative of this similarity: “Anybody who thinks this is all about drugs has his head in a bag. It’s a social movement, quintessentially romantic, the kind that recurs in times of real social crisis. The themes are always the same. A return to innocence. The invocation of an earlier authority and control. The mysteries of the blood. An itch for the transcendental, for purification. Right there you’ve got the ways that romanticism historically ends up in trouble, lends itself to authoritarianism. When the direction appears. How long do you think it’ll take for that to happen?”
The Golden Gate bridge in San Francisco, 2023.Painted Ladies in San Francisco, 2021.Castro neighborhood, San Francisco, 2021. Ocean Beach, San Francisco, 2022.
The alleged counterculture that we see today in San Francisco’s tech scene is very similar to the quintessentially romantic social movement of the hippies. The innocence manifests through the Peter Pan syndrome, the desire to never grow up and to continue playing and exploring. An itch for the transcendental is the ignition that propels so many people in the city’s tech scene to build, to create, to “change the world,” all in the name of greater good and nonconformity. And sometimes that itch is the same one that the hippies had, which gets scratched with drugs and by getting high; not necessarily to be numbed, but to transcend.
Didion makes a great remark in the essay that she had been witnessing children detach from their roots to create a community in a social vacuum, an irrefutable evidence of the society’s atomization post World War II. Though she does not state this explicitly, I believe she was never dismissive of the pacifist values that were born out of this process in the sixties. But what she does point out indirectly, through the stories of her Haight-Ashbury friends, is that these counter movements often end up not so counter after all, precisely because they seek to establish the same values many of us run away from.
For Max, Sharon, Tom, Barbara, and many of the other characters in the essay, their counterculture meant rebelling against the norms imposed by the previous decade of suburban boom, familial rigidity, and growing American corporatization. It meant trying out every drug, resisting commitment, defying suburbia gender roles, and letting everything be “groovy.” And yet, what Didion witnessed were weeping cries to find a home, carefully orchestrated trips on acid, teenagers who want to commit but cannot, and girls who dislike “earning more than $10 or $20 a week” and most of the time “keep house and bake.”
For many of us who moved to San Francisco to be part of the tech scene, counterculture meant rebelling against stability, predictability, and employability, all of which had been the norms imposed by the aftermath of the 2008 financial crisis. It meant dropping out of school instead of getting that prestigious degree, working for a startup instead of a predictable office job that “opened doors,” hanging out with “builders” instead of networking with VPs and SVPs and the C-suite, and thriving in ambiguity instead of seeking stability. And yet, a few years down the line, we find ourselves job-hopping between startups to optimize our comp packages, waiting out on the IPO to get those downpayment checks, coveting nomination-only memberships at San Francisco’s clubs for the tech elite, and having mentors at work to “cope with uncertainty.”
This similarity is why I felt so transfixed by Didion’s essay when I first read it. I think her experience helped me understand why I often felt cheated after getting to know the mainstream San Francisco scene. I expected to see defiance, maverick values, valiant impulsiveness, but in reality, most of us so far have been fairly predictable characters with a well-established, longstanding, approved narrative.
From the 1960s to the the 2010s and now to the 2020s, the society’s atomization has continuously produced us, characters who detach from their nuclei and seek to rebel against society’s norms, ending up in idolized communities that, in the end, seek to establish the same values many of us run away from. What has plagued me though, each time after reading Didion’s essay, is that I haven’t felt the same level of despondence that she did in 1967 when writing this story. And I think I finally understand why.
Didion’s implicit argument in the essay is that nuclear, familial structures are necessary to prevent further atomization and disintegration of American society. As she says, “Once we had seen these children, we could no longer overlook the vacuum, no longer pretend that the society’s atomization could be reversed. […] These were children who grew up cut loose from the web of cousins and great-aunts and family doctors and lifelong neighbors who had traditionally suggested and enforced the society’s values. They are children who have moved around a lot, San Jose, Chula Vista, here.“
But what if today the society’s atomization doesn’t need to be reversed? What if those cousins and great-aunts and family doctors and lifelong neighbors are not good for us? What if their values are oppressive and suffocating and ignorant? What if both statements can be true? I am part of a community that exists in a social vacuum, composed of people who are not actually rebelling but merely running away from the norms that suffocated them. And, it’s a community that, however faulty and paradoxical, is the right place for me—at least for now.
The reality is that all of us here have detached from our nuclei for some reason. I love my real, nuclear family. But I’ve always known that, for my own sanity and quality of life, I had to embrace society’s atomization and detach. I had to build a life in which I would find a chosen family, far away from home and far away from my nucleus. Maybe it’s selfish, but I accept this paradox now and I no longer see the mainstream San Francisco scene with suspicion. Instead, I see it with empathy. Because I am a part of it.
I think we all have to ask ourselves ultimately, if our society keeps getting atomized, and there are people who not only embrace the atomization but actually need it, then is the problem in the society’s atomization or in the society itself?
I grew up in a family that loves perfumes. Some of my earliest childhood memories are those of my mom going out with my dad to formal social events in the winter, wearing this unusual, glamorous earthy scent that I had no way of describing as a child but that vividly encapsulated my mom’s personality. Powerful and loving but also enigmatic and defiant.
My grandma, who was always dressed in bright pastel colors, wore sweet, floral, and powdery perfumes. They suited her because she herself was ebullient, theatrical, melodramatic, but also non-confrontational. And, even my dad, who isn’t one to spend time overanalyzing beauty products, allowed my mom to choose a small collection of colognes for him—often spicy, balsamic, and a bit woody—that complemented his warm but strong-willed nature.
My brother and I, as a result, got exposed to this hobby early on and learned to appreciate the art of perfume-wearing. I wasn’t aware this contextual education was happening, but as I got older and as my collection grew, my friends would point out that it was unusual I had so many perfumes, and that I had strong opinions on when and which perfume to wear. To me, of course, it wasn’t unusual because this had always been a thing in my family.
Interestingly, I only realized later that my way of deciding when to wear which perfume was very different from that of my family. While my mom, for example, can instantaneously break down each perfume into top, middle, and base notes, which somehow tell her when to wear it, my mind thinks very mathematically. When I first try on a perfume, I see an image in my head, one that I can clearly contextualize with the help of time, temperature, and weather. And then, my mind translates those features to math.
That math is something that I regularly use to choose which perfume to wear, and something that I think can bring a lot of ease to everyone’s decision-making.
The 3D Cartesian coordinate system for perfumes
For me, each perfume can be a placed in a three-dimensional Cartesian coordinate system, where the x axis represents the time of day, the y axis represents external temperature, and the z axis represents precipitation.
On the x axis, we move from day to night. On the y axis, we move from hot to cold. And, on the z axis, we move from clear to rainfall. I make the z axis simplified on this chart because I usually consider any form of precipitation, rain or snow, at the “rainfall” end of the spectrum.
Now, when we talk about Cartesian coordinate systems, we have to have some form of numbers represented on these axes to describe position.
For simplicity, I like to imagine that each axis is bounded by —1 and 1 as the terminal points. What that means is that on the x axis, just as an example, —1 would represent the late night and 1 would represent the early morning, just after sunrise. On the y axis, on the other hand, we can imagine that —1 would mean really cold weather and 1 would mean really hot weather.
Keep in mind that these terminal points on each axis are directional at best; they are not mapped to the exact magnitude of these situations in real-life scenario. In other words, when I say 1 on the y axis, I don’t think of it as boiling-level temperature, but instead as just the regular summer heatwave.
You can start to place various points in this space and identify what they would mean in real-life scenario. For instance, what would the origin O with x = 0, y = 0, z = 0 mean?
For the x axis, that would refer to twilight, that majestic moment between the day and the night. Though, in the context of perfume-wearing, I am really referring to the time just after sunset, because I could not imagine myself waking up and timing perfumes before sunrise.
For the y axis, that would refer to the perfectly moderate, mild temperature of the environment that’s neither hot nor cold, which for my body is 21°C or 70°F.
Finally, for the z axis, that would mean the classically cloudy day, when you are not able to see the sun at all but there is also no rain or snow.
Together, that would imply that the origin O would represent a moderate 21°C or 70°F cloudy day, just after sunset. Who thought that dreadful high school math would turn out to be actually applicable?
My favorite perfumes in the 3D Cartesian coordinate system
Let’s see how some of my favorite perfumes can be described in a 3D Cartesian coordinate system. I picked out the three that I think are very different in their olfactory composition and intended use: Fahrenheitby Dior, Un Jardin sur le Nil by Hermès, and Yataganby Caron.
Dior—FahrenheitHermès—Un Jardin sur le Nil Caron—Yatagan
Fahrenheit was a cologne that I discovered in 2018 and instantly fell in love with. The moment I smelled it, I knew it was a statement scent. Spicy, citrusy, woody, and leathery, with a prominent underlying note of gasoline, I instantaneously saw myself wearing it when going out at night, wearing a punk-inspired jacket, black leather boots, a sweater or a shirt with skull imagery.
So, my mind translated that imagery to: late night, crisp cold but not freezing weather, and clear skies—because I don’t do rainy nightlife. Mathematically, I therefore assign Fahrenheit to point DF in space with coordinates x = —1, y = —0.5, and z = 1.
Un Jardin sur le Nil was one that I was introduced to by my mom. I distinctly remember seeing her wear it in springtime, often when she had to attend a work lunch, when she would also sport a sophisticated, chic outfit. At the same time, the perfume with its floral and citrusy notes, mango being the most prominent one, gave her a playful edge.
As a result, my mind always associated this one with an image of a gentle but playful butterfly, flying on a bright day, in peak spring, when it’s usually clear with maybe an occasional spring rain that gives even more life to all the beautiful nature around us. Mathematically, I therefore assign Un Jardin sur le Nil to point HN in space with coordinates x = 1, y = 0.35, and z = 0.75.
Yatagan I discovered while reading an interview with the incredible Iris Apfel. She mentioned that she and her husband Karl used the same perfume by Caron, which was not very popular at the time. After some research online, and perusing many people’s reviews, I found it for a very reasonable price and decided to do a blind buy. I have always been such a fan of Iris, I figured she must know if something is good.
I remember opening the box when it came and seeing this amber, chic-looking perfume bottle with dark-red letters spelling out Yatagan (later on, I learned it was a type of Ottoman knife) in a font that very much reminded me of the Prince of Persia video games.
The first whiff gave me strong wood and forest imagery. I liked it. I imagined myself walking through a dense forest on a fall day, stepping on freshly fallen autumn leaves, damp from all the rain, which accentuated the smell of the surrounding pine trees. Mathematically, I therefore assign Yatagan to point CY in space with coordinates x = 0.25, y =— 0.5, and z = —0.8.
Each perfume that I discovered or was introduced to was therefore always followed by an initial imagery, which I could always describe using the time of day, external temperature, and precipitation. My mind would then map the components of this imagery to a three-dimensional Cartesian coordinate system.
And that’s how I ended up with a list of all my favorite perfumes with their x, y, and z coordinates. Whenever I have to decide which perfume to wear, I observe what the day is like and consult my little database of coordinates.
Using the 3D coordinate system to group perfumes by their vibe
What’s cool and fun about placing perfumes in a 3D coordinate system is that we can calculate the distance between two perfumes. This distance is what my mind sees as the difference in their “vibe”.
Because this 3D coordinate system has contextual significance, with each axis representing something that we experience with our senses, the proximity of perfumes in this space tells us that they elicit a similar… well, for lack of better word, vibe.
That distance can be calculated using the well-known equation for calculating the distance between two points in a three-dimensional Cartesian coordinate system. The larger the distance, the greater the difference in vibe.
For orientation, it’s useful to identify the smallest possible distance dmin and the greatest possible distance dmax between two points. The smallest distance is obviously zero. The greatest distance is less obvious, so we actually need to calculate it. The formula above shows that the greatest distance will happen for points P1 (1, 1, 1) and P2 (—1, —1, —1), whose distance in vibe is 3.46.
In my personal collection, as seen in the tables above, Terre d’Hermès Eau Tres Fraiche has coordinates x = 1, y = 1, z = 1. I don’t, however, have any perfumes that have x = —1, y = —1, and z = —1. The closest perfume to that theoretical point P2is Promise by Frederic Malle, with coordinates x = —1,y = —1, and z = —0.4.
Hermès—Terre d’Hermès Eau Tres FraicheFrederic Malle—Promise
These two perfumes are indeed almost the complete opposites. Their vibes are very different.
Terre d’Hermès Eau Tres Fraiche is a pleasant, zesty, inoffensive summer fragrance that’s perfectly suited for a day on a boat, wearing a light, breezy white shirt, white chinos, and beige espadrilles.
Promise is a bold, risky, and very divisive scent that I can pull off only when I am feeling incredibly confident and even a bit cheeky. It’s sexy, intoxicating, and frustrating. It juxtaposes rosy, woody, and spicy notes, which means it rebels. And rebelling means it’s not a crowd-pleaser, so I wear it only when I feel really good about myself.
It can be hard to visualize how vibe relates to mathematical distance, so let’s plot two perfumes, Fahrenheit and Un Jardin sur le Nil, that we placed previously in the 3D space and calculate the distance between them.
Looking at the chart, we can see that these two perfumes also have a decent distance between them, which means we should expect that the math would tell us the difference in their vibe is high—not 3.46 maybe, the greatest possible distance, but at least somewhat high.
2.19 is a solid distance! And that number correctly shows how I—and, I imagine, many other owners of both perfumes—feel about their opposing vibes. Fahrenheit is a gritty nightlife perfume, while Un Jardin sur le Nil is a sophisticated, garden party perfume. They are, however, brought closer together in space by being best suited for days with clear skies, which is why their distance in vibe is not as high as 3.46.
Final thoughts
Nailing down the art of knowing when and which perfume to wear is a fun and rewarding undertaking. I always enjoy learning how others choose their perfumes and how their minds analyze them.
If you are someone who is not into perfumes, but would like to know more, I recommend starting with Fragrantica. It’s like an encyclopedia of all perfumes out there, and a wonderful place to learn the major names in perfumery, to understand how others choose perfumes for their collections, and to pick up all the cool incisive vocabulary that will help you describe olfactory experiences in a far more nuanced way.
🕒 This essay is more than 5 years old (Published Jan 19, 2017).
Sustainability has become such a buzzword recently. Sustainable practices, sustainable development, sustainable world, green economy, environmentally-friendly business practices, and so on—we hear them everywhere. Obviously, all these concepts are responses to the rapidly-developing industrial world around us, which has now clearly left us in a precarious state, one that is advanced and innovative but brittle and shortsighted.
Is there actually a solution to this? Is there a path to a more sustainable future for our world? Probably, but our society will have to look beyond the typical pop-culture ideas, such as green economy, to understand that, in order to pull ourselves out of this unsustainable industrial state, we have to think about all areas of sustainable development: economy, environment, and employment.
For me, that means we have to first focus on energy and figure out a way to use it for co-optimizing these three pillars of sustainable development. Energy has tremendous potential to spur innovations and lead to economic growth, but also to hurt the environment, as has been shown throughout history. As a result, we are left with a powerful tool that is a double-edged sword and that has to be used effectively and responsibly.
I also think it’s necessary to create meaningful and rewarding employment opportunities because technological innovations will not have any significant effect if working conditions remain the same—unpredictable, insecure, and largely underpaid.
The reason I think we need to focus on these two solutions is because they tend to be more realistic than what we often hear from scholars and policy-makers. This is not to say these brilliant minds are wrong, but simply that we can still look beyond those proposals and determine what is actually doable.
So, let’s start with what they all have said so far.
Quick summary of what others have said about creating sustainable future
Many of those who advocate for transforming the unsustainable industrial state refer to green economy, a term that has undoubtedly become a buzz word in the realm of efforts devoted to sustainability. While it is true that the mainstream use of the term primarily refers to economic growth with positive environmental effects, scholars like Borel-Saladin and Turok have pointed out that green economy also serves to have a social impact, such as alleviating poverty.
Examples of such policies within the realm of green economy include providing sanitation the poor, or introduction of solar power to replace coal and generate jobs for solar manufacturers and technicians. Recommendations of broader scope for interventions necessary in green economy range from education and innovation to infrastructure and close involvement of government.
Others see green economy as a tool that is closely related to local communities, rather than large-scale systems. For instance, Jackson and Victor argue that we need to “use as little as possible in way of materials and energy,” arguing that local infrastructure and resources are paramount for achieving green economy, and that financing community-based energy is essential for green economy because most renewable sources and materials are local. They see enterprises as central agents of developing green economy through local communities, but that state actors can help this process by serving as catalysts, similarly to Borel-Saladin and Turok’s propositions.
For some, green economy is not an end goal, but rather a stepping stone to a society that sees sustainable development as means for more responsible consumption. Lorek and Spangenberg propose strong sustainable consumption strategy, which would force us to rethink how we perceive mainstream propositions of sustainable consumption, or as they call them, practices of “weak sustainable consumption.”
So, for example, instead of understanding ‘green supply chains’ as supply chains in which firms force their suppliers to disclose environmental standards, global firms should ensure that there exists a “standardized information flow along the product chain about ecological backpacks as well as social standards in the companies.”
Interestingly, the authors also advocate for precautionary principle, which suggests that radical innovations should be implemented with no expectations of them eventually coming to life, so that society can be prepared in case of their non-appearance. Either way, they agree with the previously-mentioned authors in that significant structural changes in the society and people’s way of thinking are necessary for transforming the industrial state.
While green economy implies economic growth, some authors argue that degrowth movement is necessary to achieve such transformation. Kallis, a proponent of such ideas, claims that, even though degrowth is essentially negating the mainstream concept of ‘growth,’ it is necessary to reduce society’s throughput to move toward sustainable development, even if that implies reduced GDP.
Additionally, Kallis shows that significant reformations are necessary; in fact, he argues that movement of degrowth is in accordance with model of revolutionary social change, and that as a society, we need a new political project, rather than new environmental policies in order to avoid a crisis.
Beyond the concept of green economy, some environmentalists are also proponents of more specific policies, such as reducing work hours because they think such policy can secure employment without growth and reduce consumption in the long run (Ashford and Kallis, 2013, p. 53). However, it is also clear that policies like these are not bulletproof and that they need to be accompanied by more systemic changes, such as those in taxation, accompanied by increasing working and poor people’s access to capital.
What can we learn from all these experts on sustainable growth? It boils down to—each of these policies should not be delivered in isolation; they need to be complemented by large-scale changes of the society, from more palpable aspects such as energy, infrastructure, and government, to less palpable ones such as people’s mindsets. The implicit argument in all these publications is that such radical changes are not easy or practical to achieve.
Let’s break down why.
Why policies of sustainable growth are harder than they sound
It’s important to think about the ways these policies will actually affect our society and whether they are necessarily practical.
Starting from the concept of green economy, while it is true that such approach might lead to improved environmental performance and possibly positive employment effects, one needs to consider several drawbacks. For instance, valuing ecosystem goods and services is difficult, market imperfections are serious constraints (e.g. price signals are not credible and predictable), and there are inadequate financing tools for up-front investment in green technology.
Referring again to Borel-Saladin and Turok’s analysis, even though the authors of the paper are aware of some drawbacks in “greening” the economy, it also seems that some of their suggestions lack more in-depth analysis. I mentioned earlier that one of their propositions for greening the economy is to introduce solar power instead of coal, which would generate employment opportunities for solar manufacturers and technicians.
However, this sort of ‘replacement’ innovation will also displace jobs of those who were trained to work in traditional industries, as was suggested by other scholars (Edquist et al., 2001, p. 120). Put differently, even these well-intentioned policies can have serious repercussions if they are not inspected with attention.
Similarly, when Jackson and Victor propose that green economy can be achieved by placing faith on local communities, it is important to understand that this is not always practical or easy to achieve.
They claim, for instance, that “the seeds for this new economy already exist in local, community-based social enterprise: community energy projects, local farmer’s markets, slow food cooperatives, sports clubs, libraries, community health and fitness centres, local repair and maintenance services, craft workshops, writing centres, outdoor pursuits, music and drama, yoga, martial arts, meditation, hairdressing, gardening, the restoration of parks and open spaces”.
Admittedly, the authors are aware that local communities cannot carry the entire burden of green economy, and they support these claims by arguing that governments need to be extensively involved in the process, but we need to discuss what a realistic potential of local communities in developed versus developing countries would look like.
Particularly, the authors argue that people enjoy and harbor greater fulfillment from the previously-mentioned activities than they do by spending time in “materialistic, supermarket economy in which much of [their] lives is spent.” Certainly, this can be true, but I doubt that their ideas would be implemented in developed and developing countries with equal effectiveness.
In most of developing countries, where there is already a stronger sense of community between people, the seeds of “new economy,” as Jackson and Victor call it, will be easier to sow because the mentality of people in developing countries is more likely to be complementary to the authors’ vision of community-centered economy.
Of course, I am not arguing that such transitions will not be possible in developed countries, like the United States, but they will require intensive adjustments of people’s mindsets. Those who are accustomed to supermarket-driven overconsumption and the “more is more” mentality will need to go through more significant adjustments if they are to be convinced that downscaling to community-sized economies is a satisfying outcome. This line of thought circles back to suggestions of authors like Lorek and Spangenberg, who argue that radical disruptions of societal systems are necessary if we want to implement these policies.
If these ideas seem unpractical or hard to implement within a reasonable timeframe, it might be tempting to think that Kallis’ pro-degrowth suggestions are feasible. Certainly, it makes sense to decouple development from growth given that many issues within the current industrial state have been engendered by overconsumption and materialistic desire to always have more. However, even though Kallis doesn’t explicitly state his preference for drastic strategies, such as “exit from the economy,” it does seem that he is calling for a major overhaul of capitalism by strongly arguing for degrowth movement.
Is that even practical? I don’t doubt that his approach of degrowth would mend the currently existing fractures within the society, but I would say it’s more practical to think about the ways that growth—through tools such as energy—can be used to co-optimize the three pillars of sustainable development. This would contradict the idea of radical changes previously mentioned, but I would say it might be more feasible more because overhauling capitalism via degrowth could easily result in a major fiasco.
I also briefly mentioned earlier some of the discussions that have been happening around the reduction of weekly working hours. Instituting a shorter workweek is not guaranteed to have a positive impact on employment and the environment in all contexts. As Ashford and Kallis argue, it is more beneficial to reduce the workweek to four days without changes in weekly wages (thus increasing hourly wages), otherwise the underlying implication would be that workers’ wages are sacrificed for the sake of environment.
Second, even though it seems that shorter workweek might decrease consumption as people would spend more time at home, thus performing leisure activities, it is also possible that leisure would lead to other forms of consumption. That is, the effects of a reduced workweek are strictly positive only if we assume that workers will spend their free time in an environmentally-friendly way. A bold assumption, in my opinion.
When all is said and (un)done, then what are we even left with? Is there a way forward?
Ways to transform the unsustainable industrial state
Comparing all the proposed arguments, it seems as if there is no substantial solution for transforming the unsustainable industrial state.
On the one hand, some scholars think that specific policies, such as greening the economy, opting for degrowth, or reducing the workweek, will be sufficient to provide foundations for sustainable development. On the other hand, other scholars argue that isolated policies are not enough and the we need more radical transformations in order to achieve these goals.
Such significant changes, at the same time, require time and unprecedented effort that it seems more practical to rely on a carefully chosen set of specific policies. It certainly seems like a vicious circle, but I think we can achieve a middle ground by:
focusing on effective use of energy and
creating meaningful and rewarding employment
These two approaches can provide specific suggestions that could introduce significant changes and begin to drive the reformation of the industrial state.
Energy as a tool for co-optimizing the three pillars of sustainable development
Throughout this analysis, I hope it has become clear that, among the commonly mentioned topics such as innovation and government intervention, energy plays a key role in thinking about sustainable development. It has been a central factor in spurring innovations in the past, so it seems more than fitting to re-orient the discussion of sustainable development to energy and its use in the near future.
Most importantly, the paradoxical implications of energy use, which is that it can positively impact economic growth and negatively affect environmental conditions, leads to an important question—how can energy be used to have positive impacts on all three pillars of sustainable development: economy, environment, and employment? The answer will vary slightly between developed and developing countries, but in both cases, innovation, as well as sustainable production and consumption, are essential for effective use of energy.
Innovation of safer products and processes can greatly impact how energy is used, but one must understand that not all innovations will have positive impacts on all three pillars of sustainable development. As scholars have shown previously, even if we assume that product and process innovations will be environmentally safer and beneficial to the economy, they do not always have a positive impact on employment.
Particularly, product innovations generate employment if they do not substitute old products and if they do not become process innovations in the long run, while process innovations generally reduce number of jobs, even if they increase other aspects of employment, such as productivity. If one thinks about the ways these implications affect potential use of energy for co-optimizing all three pillars of sustainable development, it becomes clear that some compromise is necessary.
For instance, Borel-Saladin and Turok have used example of solar power replacing coal, which would not only be an environmentally friendlier use of energy for advancing economy, but would also generate employment for solar manufacturers and technicians. This is a great example for applying the analysis proposed by Edquist et al. because it represents, essentially, both a product and process innovation.
Certainly, by substituting coal with solar power, new employment opportunities will arise for workers such as solar manufacturers and technicians as the markets will require knowledge and skills for new types of products, but as the new technologies will be accompanied by new processes, it is important to think about what will happen to those workers who have been trained to work in industries with coal.
If this substitution happens with no complementary adjustments, the workers who cannot keep up with new products and processes are likely to lose jobs as more traditional industries are displaced. The popularized idea that coal workers will simply “adjust” and learn new skills is a very rosy view of the world—we need to be more empathetic to those who will experience that change.
This means that energy-based innovations should be complemented with education and training of those who are employed by industries that are going to become displaced by rising innovations. Of course, adjustments like these would require a substantial reformation of some societies (e.g. high-quality education should be accessible to everyone), but it is paramount that innovations do not lead to massive creation of jobs that require highly specialized skills, only attainable by those who have the means to acquire them.
Referring again to the analysis of Edquist et al., I doubt that it’s realistic for institutions, such as the government, to encourage only those types of product innovations that do not substitute older products, which is why it is important to make a compromise when less beneficial energy-oriented product and process innovations occur. In that sense, governments should encourage innovations that are complemented by education and training of those workers who have been trained in more traditional industries. This way, innovative use of energy would advance the economy in an environmentally friendly way while ensuring that jobs are not lost in the process.
Sustainable production and consumption are also necessary if energy is to be used responsibly for transforming the industrial state. In addition to moving to cleaner and safer technologies and processes, it is important that less energy is used in production and provision of products and services that require less energy in their operation and use. Although it might seem unclear how this can co-optimize the three pillars of sustainable development, sustainable production and consumption ensures that economy can advance (as consumption will decrease) in an environmentally friendly way (less energy is used), while increasing long-term security for workers due to a positive feedback loop promoted by advancing economy, resulting from more responsible production and consumption.
These ideas should be implemented in both developed and developing countries, but in countries like China and India, there needs to be a greater compromise when thinking about the repercussions of improper energy use. As Lorek and Spangenberg have noted in their paper, these countries are justifying their growing emissions by claiming that their ever-growing populations cannot be complemented with reduced environmental impact. The authors do not agree with this notion, and instead suggest that the right to growth, and therefore increased emissions, should not be granted to any country, but to groups of people in poverty.
For effective and responsible use of energy, this means that, as a society, we might have to compromise and allow high-poverty areas to temporarily use forms of energy that might lead to negative environmental effects, which would not co-optimize the three pillars of sustainable development. While this might not be the most ideal way of transforming the industrial state, it might be more practical until governments can redistribute wealth from affluent to high-poverty areas, allowing them to also institute energy-oriented innovations.
Meaningful and rewarding work opportunities as tools for ensuring workers’ enduring satisfaction
These innovations will not have a significant impact if the new employment opportunities continue the same trend of unrewarding and underpaid jobs that are harrowing many of today’s workers. Solutions, such as financial compensations above the minimum wage, seem clear and unambiguous, but it can be hard to assess what “meaningful” job opportunities mean.
Meaningful work implies at least some sort of personal fulfillment; that is, while it is unrealistic to expect that all workers will do what they love, it is important that workers can perceive their contribution to the society as significant. One way to achieve this is to offer product services wherever and whenever possible.
Using the example of solar power again, one can imagine a hypothetical future scenario in which solar power systems can be purchased by consumers for delivery and do-it-yourself assembly system, which would help companies cut costs by not employing technicians who would help assemble such power systems.
This line of thought would follow the widespread, profit-maximizing mentality of the today’s world—just think of IKEA—but if we are to transform the industrial state, the companies should orient their business model toward sustainability and employ technicians instead of allowing consumers to assemble these systems on their own. Even if the consumers are capable of doing so.
First, while this approach might increase labor costs for the company, it would generate job opportunities for workers who are trained as technicians. Second, this approach would generate meaningful work for technicians as they would be able to apply their hands-on skills and customize the solar power systems per customers’ desires, which might not be necessarily possible in a simplified, do-it-yourself system. In turn, this would provide the workers with a sense of nontrivial impact, which is, I would argue, the basis of meaningful work.
Bringing it all together
My main argument here is that no policy or suggestion can be a guaranteed solution for transforming the unsustainable industrial state. Even though some propositions might seem effective, if they are implemented in isolation or without in-depth, pre-emptive analysis, they might engender unwanted consequences.
Likewise, calls for radical transformations of society are theoretically more effective, but it is not always clear whether they can be achieved within a reasonable timeframe. Despite this odd conundrum, we can start to transform the industrial state by focusing on two areas: using energy to co-optimize the three pillars of sustainable development and creating meaningful, rewarding work opportunities.
They might not be courageous or radical ideas, but I would think they are at least somewhat more practical.
🕒 This essay is more than 10 years old (Published Oct 13, 2015).
Meds, clothes, and intellectual property. Probably the last three words you would imagine together in a sentence. But, as preposterous as this may sound, I would say that the fashion industry can learn a lot from the pharmaceutical industry. These two seemingly disparate worlds actually have something in common—a market failure: unstable and undefined intellectual property (IP) rights.
To disentangle this paradox, we need to dissect the IP issues in both industries, and then draw a two-fold comparison to show how current IP challenges in the pharmaceutical industry can actually be analyzed to build foundations for IP policy-making in fashion.
I definitely do not aim to offer sound policy for rectifying these issues in both industries, but I do think it’s valuable to point out why a very notorious industry, known for its highly competitive and sometimes unscrupulous regulatory processes, should be seen as a template for protecting the work of fashion designers, people who unfortunately often live at the blurry intersection of art and commerce.
Unstable and Undefined Intellectual Property Rights in Fashion
Why would one even think about this issue in the fashion industry? Well, more than anything, there is an economic importance. A 2008 report from the U.S. Census Bureau estimated $217 billion in annual sales from clothing and accessories, while shoes generated $27 billion through retail. A more recent report in The Legal Intelligencer cites $300 billion in sales as the industry’s annual revenue, which shows that the industry contributes a significant amount of revenue to the US economy and that the sales are likely to continue growing.
Despite its economic importance, at the same time, the industry lacks well-defined and established laws and policies that protect intellectual property in US-based fashion, which can easily allow others to copy ideas in fashion without significant repercussions. This, however, is not the case with non-US branches of the fashion industry, such as the European one, where fashion designs can receive up to twenty-five years of protection.
Recently, there have been few attempts at solving the issue of intellectual property in fashion through copyright, but they have all, unfortunately, failed. For instance, The Design Piracy Prohibition Act of 2006 never passed from Congress to the next steps approval, and the Innovative Design Protection and Piracy Prevention Act of 2010 was placed on the Senate Legislative Calendar at the end of 2010, but has not since moved forward. The Innovative Design Protection Act of 2012, which would have allowed protection of fashion designs up to three years, was introduced at the end of 2012 but was never passed.
While fashion designs themselves cannot be protected by copyright, some aspects—unique prints, patterns, and color arrangements—can be protected “only if, and only to the extent that, such design incorporates pictorial, graphic, or sculptural features that can be identified separately from, and are capable of existing independently of, the utilitarian aspects of the article,” according to the U.S. Copyright Act.
Of course, one can immediately see how this vague rule can be easily surpassed to copy an idea without creating ground for litigation. For example, in a hypothetical situation, if a designer uses a triangular color-blocking pattern—a pattern that is not novel—to create a novel line of color-blocked scarves, which no one has presumably thought of before, the designer will not be able to protect neither the color-blocking scarves nor the design as the triangular color-blocking pattern might not be identified separately from the utilitarian aspects of the scarves. An obvious global example is Burberry’s popular cashmere scarf in heritage check, whose characteristic pattern can be easily replicated with only slight modifications and used to sell unbranded, Burberry-like scarves on popular commerce websites, such as Amazon.
Designers can partially rectify these issues by turning to trade dress, a part of trademark law governed by the Lanham Act, or design patents, which are available under the U.S. Patent Act. Trade dress allows designers to protect their brand names and logos, as well as the visual characteristics of a product if they denote the source of the product to consumers and if they are not functional, which effectively excludes apparel since apparel designs are considered to be functional.
Design patents, on the other hand, will protect the look of a design and ornamentation as long as “it is novel, nonfunctional, and nonobvious to a designer of ordinary skill in the art.” Once again, it is clear that there is great degree of freedom in interpreting this statement, essentially leaving designers without a guaranteed safety net. On top of that, design patents are generally not applicable to apparel because of its functionality.
You still might be thinking: why does anyone care? After all, if a person is a loyal Oscar de la Renta customer, for instance, it is reasonable to assume that they would not bother looking into non-original garment that’s sold for less. While the assumption is valid, as it was shown in BBC’s popular 2007 documentary The Secret World of Haute Couture, there is only a very small and exclusive global community that religiously attends designers’ publicly-inaccessible runway shows and that purchases these exclusive designer garments.
The rest—in fact, the majority—of the population has easy access to non-original products, which can be classified into: counterfeits, knockoffs, and generics.
Counterfeits are those products that are produced with the intent of being sold as originals, while knockoffs are derivative products that are not meant to be sold on the black market as originals, but they have been designed by adopting one or many ideas from an original product.
For instance, using again the example of Burberry’s popular cashmere scarf, a counterfeit would be a replica of the scarf that’s illegally marketed and sold as the original one, while a knockoff would be a Burberry-like scarf that’s sold on Amazon under a different name.
Scarf, Burberry originalScarf, Burberry knockoff
It is therefore clear how a counterfeit might negatively impact revenues and profits of a branded fashion house, but one might naturally wonder how high-end fashion knockoffs can have a negative impact if they are marketed under a completely different seller, such as a retail-clothing company.
As Ferrill and Tanhenco describe in their paper, knockoffs are not prohibited by any U.S. law—while counterfeits are regulated by trademark law and the Lanham Act—which means they are even more dangerous for the world of fashion, because they hurt designers both economically and creatively.
Particularly, the knockoffs can hurt designers economically because those customers who are willing to pay enough for a high-end fashion product, but also do not mind purchasing a knockoff, will buy the knockoff at a lower price, which will effectively decrease designers’ revenues. As a consequence, knockoffs will also negatively affect the creative side of designers’ works; without protection over their creations, the designers will lose incentive to invest in developing novel designs.
One can think of a situation where this would be particularly relevant—if the creation of novel clothing line requires high costs of production due to a specific material or intricate pattern-making process, a designer might not have the incentive to incur high costs if the product can be easily copied and sold as a knockoff in retail-clothing stores, such as H&M or Zara.
Finally, one can think about the long-term negative impacts in this industry caused by the lack of strong protection over intellectual property. While counterfeits and knockoffs represent immediate threats to designers and high-end fashion houses, it is important to understand that there is a third type of threat—not yet presented and analyzed in the literature—which comes from products that have lost their patentability potential due to their overuse over an extended period of time.
I call these fashion items generics as they are closely comparable to generic drugs in the pharmaceutical industry, drugs marketed under their chemical name that are identical to corresponding brand drugs in qualities such as dose, strength, and efficacy. It is easy to see how generics can develop in fashion if there is no firm protection over a new design and product. For example, if a designer creates a trapezoid-neck t-shirt, which has presumably not yet appeared in the world of fashion, and is unable to protect this unique design, the t-shirt will be able to get copied as a knockoff in the near future and finally become a generic in the long run.
Put differently, after an extended period of overuse in the retail-clothing industry (assuming that the product prevails throughout multiple seasons), the concept of trapezoid-neck will become obvious and not new, effectively annulling the patentability of the original design idea.
Generics do not economically impact designers and high-end houses in any different way than knockoffs do, because they stem from knockoffs, but they can hurt them creatively in a significantly more damaging way. If designers fear that their unique designs will become generic due to uncontrollable production and imitation, they will not have any incentive to reveal their ideas to the world of fashion. This, in turn, can impact the overall economy by not bringing new high-end products to the market and it can negatively affect the creativeness of the industry—simply because the designers’ ideas might be left concealed and unrealized.
So, we can identify the reason why the fashion market experiences economic and creative instability: because there isn’t a cohesive U.S. law that would establish relevant policies for rectifying the issue of intellectual property in fashion. How would we go about fixing this? This is when it is worth looking at a similar problem in the pharmaceutical industry to draw a parallel and and learn whether any IP principles from pharma can be applied to fashion.
The Issue of Intellectual Property Rights in the Pharmaceutical Industry
The pharmaceutical industry, despite its seemingly irrelevant connection to the fashion industry, faces a similar problem. Specifically, as the result of the industry’s prominent focus on research and development, especially in the field of drug development, patents are viewed as valuable assets and reliable representatives of a company’s standing in the market.
As Grabowski describes in his paper, it takes several hundred million dollars to discover, develop, gain approval, and finally send a new drug to the market. The costs of R&D are significantly high because most new drug candidates never actually reach the market due to various reasons, such as toxicity, manufacturing difficulties, and economic and competitive factors.
Even more importantly, if there is no patent protection over a new drug, imitators can easily duplicate the drug for very low costs. The costs of investing in research and developing new drugs can therefore often seem daunting to pharma executives who are aware that, especially in the field of drug development, rapid reverse engineering can lead to easy imitation of any compound.
Given that patents are currently used in the pharmaceutical industry to address these problems, one might wonder whether there is truly any similarity between fashion and pharma. It is certainly true that patenting has already been in use to protect new drugs during development, but the real issue is that pharmaceutical companies often do not have bandwidth to maintain a dedicated IP team that can strategize and help the pharmaceutical company extract meaningful profit from the generated IP.
Furthermore, any drug-producing pharmaceutical company also has to draw a distinct line between original products and those that enter the market and negatively affect company’s revenues, such as counterfeit drugs, derivative drugs, and low-cost imitations more commonly known as generics, which enter the market after the original drug’s patent has expired. These products are analogous to counterfeits, knockoffs, and generics in fashion as is evident from the following analysis.
Counterfeit drugs are fake medicines that are sold illegally under the same brand name of the original drug. As FDA describes them, they can contain the same ingredients as the original drug—although they might have no active ingredient at all—but the ingredients are incorporated at the wrong dose. One can see how they are analogous to counterfeits in the fashion industry, which are also marketed as high-end products but they usually lack in the quality or intricacy of the original design. Certainly, counterfeit drugs negatively impact the pharmaceutical industry because they can directly affect the company’s profits if they succeed at entering the market illegally and deceiving the customers.
While the term “knockoff” is not officially used in the pharmaceutical industry, and is often interchangeable with the term “counterfeit”, I use the term knockoff drugs to refer to those products that are derivatives of the original product with slight changes. For instance, a drug can be a knockoff if it performs a similar function as the original brand-name drug but has some properties changed, such as its functional groups, thermodynamics- or kinetics-based properties.
The damaging impact of these types of drugs is clear—they are not directly infringing on IP rights of the brand-name drug’s company because the knockoff drugs have different properties, but they are effectively a variation on a theme and, as such, exploit the high costs incurred by the brand-name drug’s company for R&D. Just like the knockoffs in fashion industry, which hurt designers both economically and creatively, knockoff drugs can hurt the pharmaceutical industry if companies do not ensure that their patents protect a broad spectrum of intellectual properties.
Finally, generic drugs are identical copies of brand-name drugs that enter the market after the original drug’s patent has expired. According to the FDA, they are important options that allow greater access to health care for all Americans. For the pharmaceutical industry, they can be perceived as the least threatening option because they cannot legally enter the market before the expiration of the original drug’s patent; in other words, there is an extended period of time for the brand-name drug to generate revenue before generics enter and become competition.
However, one can see how generics would be a problem for pharma if there was no clear patenting system available, which is currently the case with fashion industry. Generic drugs would immediately enter the market and, after their use over an extended period of time, annul the aspects of novelty and non-obviousness for the original brand-name drug and therefore diminish its patentability.
Sadly, this is indeed currently an issue in the fashion industry, where knockoffs can immediately lead to development of generics without leaving any time period for the high-end products to succeed at generating maximum revenue.
In her paper, Vinita Radhakrishnan proposes two strategies of strategically protecting intellectual property in the pharmaceutical industry that can rectify the previously-mentioned issues in pharma and maximize profit, and that are relevant to the examples of counterfeits and knockoffs in the fashion industry.
She proposes the use of trade secret protection for those products that are difficult to reverse-engineer or those that have value because of their secret nature. She uses an example of a biomarker that can assess the efficiency of a drug three times faster than a regular compound, and as such, can be used to optimize the company’s development, which means it should not be accessible to the public.
If placed under trade secret, the biomarker will be protected as long as the secrecy is maintained; in other words, the protection is not limited by term. On the other hand, if the biomarker is patented, the company will likely run into difficulties because the use of biomarker by another firm is difficult to audit. It is worth looking at how Radhakrishnan’s advice can be used to address the issue of counterfeit drugs.
As most drugs are easily reverse-engineered, trade secret protection could be used for those compounds that a pharmaceutical company uses as a crucial part of its drug development. As a critical component of the drug development process that is protected under secrecy, this compound will ensure the uniqueness of the developed drug so that its function cannot be copied easily. While this likely will not prevent all counterfeit sellers from distributing fake drugs to the market, it can certainly make the process harder because the difference between the original drug and counterfeit drug’s functionalities will be augmented with trade secret protection.
Conversely, patenting will work best for those products that can be easily reverse-engineered, such as a drug molecule or synthesis process. As Radhakrishnan describes in her paper, patenting does provide strong protection under law, but it comes at the expense of limited protection term, high protection and maintenance costs, as well as many exceptions to the definition of patentable matter. More importantly—and the most relevant aspect of patenting in pharma to fashion—is the drafting of specification.
Radhakrishnan suggests that the drafter of a patent should ensure that the patent is drafted to maximize the broadest possible scope while not infringing on neighboring patents. Her suggestion addressing this particular aspect of IP protection is crucial for knockoff drugs: the broader the scope of the patent is, the harder it will be to develop a derivative of the drug that’s not infringing on IP rights of the brand-name drug’s company.
It is important to note that patenting does not always benefit the entire industry. Another problematic aspect of biological research, which is an indispensable component of both the pharmaceutical and biotech industry, is the issue of anti-commons. As Michael Heller and Rebecca Eisenberg noted, too many IP rights in the upstream portion of R&D can hinder the development of downstream R&D.
According to their findings, product-developing companies have to incur high transaction costs to identify and clear rights, while academic scientists—unlike commercial scientists, who face similar problems to those of product-developing firms—seldom face patent enforcement, even when they ignore them. However, scientists in both the commercial and academic settings have trouble gaining access to materials and data that they can’t easily replicate in their own laboratories.
One can see how this essentially leads to a paradoxical problem: too many patents and high transaction costs can weaken the enforcement of patent rights, thereby increasing the possibility of unauthorized access and use, which effectively decreases the risk of anti-commons. At the same time, if the owners of the patent successfully exclude others from gaining access to their materials, users will have to incur high costs to gain access to the materials, thereby decreasing the possibility of unauthorized access, which increases the risk of anti-commons.
In their paper on intellectual commons and property in synthetic biology, Oye and Wellhausen succinctly summarize this issue through a four-quadrant graph that distinguishes private from public ownership, and clearly defined rights from ambiguously defined rights. As they describe, synthetic biologists mostly agree that processes such as protocols and design methods should fall within public ownership with clearly defined rights, while commercializable devices should also have clearly defined rights, but should fall within private ownership. They all agree that ambiguously defined rights, both within public and private ownership, serve as a threat and danger to development in the field.
Looking at all these issues together, it’s clear the pharmaceutical and fashion industry face similar problems—despite the differences between their internal operations and overall purpose. While the fashion industry’s operations focus primarily on design and transformation of raw materials, such as cotton and fossil fuels, into functional products, pharmaceutical industry places heavy emphasis on health-oriented research and drug development. Nevertheless, both industries have to combat the challenge of intellectual property rights.
Applying lessons from pharma to development of IP policies in fashion
The two most promising solutions to designers are design patents and trade dress, but as it was mentioned earlier, they are not applicable to apparel due to its functionality. Therefore, there needs to be a clearer set of policies that can protect apparel as well since, given the statistics shown earlier, apparel generates a greater percentage of revenue for the U.S. fashion industry. Using the issues and suggestions presented in the section about IP issues in pharma, one can think about potential solutions to the IP problems in fashion. My goal here is to provide suggestions for mitigating the negative effects of counterfeits, knockoffs, and generics in the fashion industry.
While counterfeits are regulated by trademark law, designers have no protection over their design process, which means that they have to keep their designs in complete secrecy until the designs are presented in runway shows or until they are worn by celebrities in public. Even so, sellers can still replicate the product and its design process because it would likely be difficult to prove that a person was replicating a fashion garment for the purpose of selling it on the black market.
Using a concept similar to trade secret, presented by Radhakrishnan in the case of pharma, policy makers should develop a law that protects the entire design technique and treats it as a secretive process. Because the design process is not similar to a synthesis process, and would therefore be difficult to reverse-engineer, it should be placed under an act that fully protects it without any time limit.
This way, if designers are using unique materials or an intricate production process, they can protect the “upstream” levels of their design technique, which would likely discourage others from attempting to obtain the same materials and knowledge about the design process. Of course, an illegal seller could still find a way to try replicating the garment, but by having a law that protects the design technique and process, designers would make this endeavor far more difficult.
Developing policies for knockoffs is a lot trickier. Drawing a parallel to the pharmaceutical industry, knockoffs can be seen as derivative products that are easy to reverse-engineer: because they won’t be exact replicates of a designer’s original garment, they are quite easy to develop through imitation. A knockoff designer simply needs to copy a single visual idea from the original designer and find a way to subtly incorporate it into a mass-production, ready-to-wear product.
It makes sense why designers would want to have patents over their final products in this case, but that doesn’t mean they should be granted these rights easily. To understand why this might be the case, it is worth looking at two examples of high-end fashion where patenting would not have benefited the fashion industry.
In a recent famous case ChristianLouboutin v. Yves Saint Laurent (2012), Christian Louboutin sued Yves Saint Laurent for selling all-red shoes, which—according to Louboutin—looked similar to Louboutin’s women shoes with red soles and different colored tops. While the court ruled that Louboutin can have trademark protection over shoes with red soles, it did not grant Louboutin permission to claim trademark over monochromatically red shoes.
Case 1: Christian Louboutin ShoeCase 2: Yves Saint Laurent, Mondrian dress
This exemplary case highlights how the concept of anti-commons could also become present in fashion if patenting was used more freely. More precisely, because fashion is a form of art, it is hard to distinguish what deserves to be patented, as patenting art forms can easily exclude others from using them in a way that’s originally intended to be non-restricting.
Simply put, it is reasonable that the court didn’t give permission to Louboutin to have trademark protection over monochromatically red shoes because that would essentially imply that Louboutin could sue any designer in the future who incorporates color red in their shoes. If designers are allowed to place strong protection over their final products, it can hinder other designers’ creativity and prevent progression of fashion.
Another example is Yves Saint Laurent’s famous “Mondrian” day dress, a wool jersey in color blocks of white, red, blue, black, and yellow, which was Saint Laurent’s adaptation of Piet Mondrian’s famous painting “Composition II in Red, Blue, and Yellow.” The dress has become one of the most important pieces in the history of fashion, yet—even though it is a high-end fashion product—the dress is a knockoff of another artist’s work.
It is unlikely that fashion connoisseurs today would consider this to be a non-inventive knockoff, but the dress undeniably represents an issue that the fashion industry is so fervently trying to combat today. So, once again, placing a patent over any design or visual feature to simply prevent creation of knockoffs would certainly hinder progression of fashion, leading to a “fashion anti-commons” issue, which is why it makes sense to have exceptions when defining patentable matter in this field of work as well.
With that in mind, what is the right way to approach this issue and still preserve designers’ intellectual rights? One way to develop a set of potential solutions is to use the framework presented in Oye and Wellhausen’s paper.
Currently, the U.S. fashion industry is having problems in the territory of intellectual property rights because the policies are based in the two quadrants of ambiguously defined rights. Policy-makers need to establish laws that can move fashion-oriented policies into the other two quadrants, by clearly defining rights for designers so that it’s transparent what can be a designer’s private ownership and what should be “intellectual commons.”
Of course, there are many ways to categorize aspects of fashion design into these categories as these decisions would be highly subjective. I propose one potential solution: categorizing these aspects in terms of structural and visual non-obviousness and innovativeness. The goal of the following policy template would be to provide designers with a firm set of rules, which would not allow for any ambiguity when deciding whether one should claim patents over their creations.
If the structural design of any fashion product, from simple accessories to apparel items, is non-obvious, and has therefore never been used before, the designer should be allowed to claim patent over such creation. For instance, in a hypothetical situation, if a designer has created trapezoid-neckline T-shirts, which therefore showcase a structural feature that has presumably not been used in fashion before, the designer should be able to protect the trapezoid-neckline structure so that this feature cannot be used by other high-end or retail-store designers who would want to mimic the product. Therefore, the designer would have private ownership over the product with clearly defined rights.
On the other hand, visual features of the product that do not in any way represent a designer’s official logo or any symbols that are officially recognized as the designer’s trademark visual features should not be patentable. In another hypothetical situation, if a designer created a line of simple dresses—which are not structurally non-obvious—that showcase an arrangement of colors gold, black, and white in a non-innovative way (for instance, a simple juxtaposition of colors along the dress), the designer should not be able to claim a patent over this dress. This way, visual characteristics, such as common patterns and colors, would be placed in the quadrant of public ownership and would constitute “intellectual fashion commons,” those aspects of design that can be used freely by all designers.
One exception to the latter rule would be the use of visually non-obvious and innovative patterns. For example, if a designer created a line of dresses that featured a specific tri-color arrangement on each dress, such that the first color was contained in an all-over triangular form, the second color was contained in an all-over circular form, with the rest of the dress featuring the third color, the designer should be able to claim patent over this specific visual arrangement, but not over the entire dress and the colors used on the dress.
Applying this rule to the Louboutin vs. Saint Laurent case, Louboutin would be able to claim patent over shoes whose soles have one color that’s different from the color of the rest of the shoe. This would prevent other designers from painting the soles of their shoes with one color, but it would still allow them to use the color for changing other features of the shoe. Therefore, this exception can be viewed as sitting in the area of clearly-defined rights, but halfway between private and public ownership because it will require a professional’s objective assessment of the visual feature’s non-obviousness and innovativeness.
Finally, as generics develop from knockoffs after the knockoffs have been used in fashion for an extended period of time, clearer policies that address knockoffs will consequently hinder development of generics. If designers are able to place strong protection over their designs and ideas, there will be an extended period of time in which high-end fashion products can generate maximum revenue before protection expires and before knockoffs enter the market and persist as generics in the long run.
Now, in the context of fashion at least, some of high-end products might survive only throughout few seasons. As a result, generics might never enter and prevail in the market if the “seasonality” of the high-end product expires before the protection over the product does. This, of course, can be seen as an additional benefit in the case of generics.
All this could certainly impose more creative protection, but I would be remiss if I didn’t acknowledge the drawbacks of my proposed solutions. While these suggestions might lead to firmer and more helpful policies for fashion designers, they would inevitably be unfair to those designers and ideas that warrant special exceptions. This is particularly true because fashion is a form of art, and developing policies for this field of work requires subjective—and therefore not always fair—decisions.
