Your Sales Team Got More Productive. What Happens to That Capacity?
I keep having the same conversation. A company rolls out AI across its sales organization. Revenue per rep jumps. Pipeline velocity doubles. The team is closing more, faster, with better targeting than anyone thought possible.
Then the board meets. Headcount gets cut.
The people who helped deliver that quarter now face cuts. The company has gained capacity and chosen to remove it.
This is the equivalent of giving your sales force cars instead of horses, watching them cover ten times the territory, and then firing them because you only needed the territory you already had. More than 50,000 workers lost their jobs to AI-cited layoffs in 2025 [2]; Block cut 40% of its workforce, Chegg 45%. The press releases read like victory laps — “growing capability of AI tools” as the justification for shrinking the organization those tools were supposed to empower.
The logic sounds clean. AI makes us more efficient, so we need fewer people. But it rests on an assumption most leaders never examine: that the amount of work to be done is fixed. That the market is what it is. That doing the same thing cheaper is the whole opportunity.
That assumption has been wrong for 160 years. It has a name. And the economist who identified it would recognize exactly what is happening in boardrooms right now.
Steel, Skyscrapers, and the Expansion Pattern
In 1865, William Stanley Jevons identified the paradox that bears his name [1]. Steam engines were getting dramatically more efficient — burning less coal per unit of work — so Britain should consume less coal. Jevons found the opposite. Efficiency made coal-powered applications viable in industries that had never used steam before. Total coal consumption didn’t fall. It tripled.
The pattern has surprised people for 160 years anyway.
The Bessemer process dropped steel production cost by 90% [4]. Producers did not spend the savings on the same amount of cheaper steel. They built skyscrapers — the Home Insurance Building in Chicago, 1885, the first steel-frame tower, launching a category of construction the cost of iron had made unthinkable. They laid continental railroads: 35,000 miles of track became 167,000 in twenty-five years. They strung suspension bridges across rivers that had never been spanned. Entire categories of building that had been physically impossible became economically inevitable.
Computing followed the same arc. When processing cost collapsed, we didn’t compute the same things cheaper — we invented the internet, then mobile apps, then social media, then cloud computing, each one “new demand” that existed only because execution cost had dropped below the threshold where someone could justify building it.
You can watch the loop running inside AI itself. As models get cheaper to run, Microsoft reported an AI revenue run rate of $13 billion in its FY25 Q2, up 175% year over year [12] — the opposite of the “we’ll need less compute now” story you’d expect if efficiency were a savings event. The companies closest to the cost curve are spending and selling more, because they understand what falling unit cost does to a market.
The same regularity shows up where you’d least expect it: irrigation efficiency raises total water use as farmers irrigate more land, and added highway lanes generate more traffic as commuters drive farther [9]. When the cost of doing something drops far enough, people find dramatically more things to do — a pattern as dependable as gravity, and about as popular with executives who’ve already committed to a layoff strategy.
AI is the next iteration. Most companies are responding the way every wrong-side-of-history company has: doing the same work with fewer people, instead of asking what new work just became possible.
Will Anybody Drink More Coke?
Coca-Cola faced the most literal version of the fixed-pie fallacy a consumer business can encounter. By the early 2000s, per-capita soda consumption in the United States had peaked and was declining. Health-conscious consumers were walking away from carbonated sugar water. The boardroom math was straightforward: nobody will drink more Coke.
They were right. Coke volume kept sliding.
They drank Vitaminwater, Monster, Smartwater, Honest Tea, and BODYARMOR — all delivered on the same trucks.
Coca-Cola expanded from a single product category to over 200 brands spanning water, tea, coffee, sports drinks, energy drinks, and plant-based beverages [5]. The $4.1 billion Glaceau acquisition and the $2.15 billion Monster stake were not defensive moves but bets that the pie was poorly defined, not actually fixed. Global carbonated soft drink volume has been flat or declining since 2005; Coca-Cola’s revenue grew anyway, because the company stopped treating its distribution infrastructure as a fixed-cost base to protect and started treating it as a platform for expansion.
The argument has a limit. Some industries face genuinely inelastic demand. A hospital cannot admit more patients because AI made diagnostics cheaper; a utility cannot sell more electricity because AI optimized the grid. The rebound literature backs this up honestly: a review in Energy Policy found rebound effects are usually significant but less than 100% [15] — efficiency does often reduce total resource use, just by less than the naive calculation suggests. An NBER survey warns that automation can lift productivity while reducing labor demand in the affected tasks [16]. The expansion is not automatic.
That qualification matters. But the hospital that adds remote chronic-care monitoring, preventive outreach, and post-discharge follow-up — programs previously too labor-intensive to staff — has expanded its definition of what a hospital does. The law firm that uses AI research to serve clients who previously couldn’t afford counsel hasn’t done cheaper law; it has expanded the market for legal services. The “fixed” demand was fixed only within the old category definition. Every physical business thinking “our market is saturated” is making the Coca-Cola mistake circa 1999: confusing the saturation of one product category with the saturation of the customer relationship. The answer to “saturated for what?” keeps changing, if you’re willing to ask it.
Klarna Cut Capacity. Affirm Expanded Its Market.
The cleanest test of the fixed-pie fallacy is happening in real time, inside a single industry. Klarna and Affirm both operate in buy-now-pay-later. Both had access to the same AI capabilities. They made opposite bets.
Klarna cut headcount from 5,500 to roughly 3,000. AI handled two-thirds of customer interactions. Revenue per employee tripled to $1.24 million. CEO Sebastian Siemiatkowski championed the cuts publicly, positioning efficiency as the model for the industry — a bold thesis timed to the company’s US IPO narrative. The metrics looked brilliant on the slide deck. They showed efficiency; customer satisfaction would reveal what those metrics missed.
For a moment, it looked like the future. Then customer satisfaction cratered.
Siemiatkowski publicly admitted “we went too far.” The company began rehiring the people it had fired months earlier — at premium rates, with institutional knowledge already destroyed. Cut the people providing judgment and service; celebrate the savings; then pay more to restore the capability. It is the predictable outcome of treating human judgment as a cost line rather than a capacity asset.
Affirm, meanwhile, grew revenue 46% year-over-year — the fastest in the BNPL sector — by investing in merchant partnerships, new financial products, and market expansion. Early European data from a CEPR study of 12,000 firms found companies adopting AI saw productivity rise by roughly 4% without reducing headcount [7]. The efficiency gains went into expansion, not contraction.
The claim needs to be precise. Jobs change. Roles restructure. Automation has always done this. The question is not whether the work shifts — it will — but whether the firm-level response is contraction or expansion: whether you treat freed capacity as a cost to eliminate or a resource to redeploy into the markets that just became reachable. History has a clear preference, and the companies ignoring it are the ones now quietly rehiring at premium rates.
The Ideas You Killed Because You Couldn’t Afford Them
Every organization I work with has this graveyard. Product concepts that got through the first three rounds of review, then died in the budget meeting. Market expansions everyone agreed made strategic sense, until someone ran the staffing numbers. Customer segments too small to serve profitably at the current cost structure. These ideas weren’t bad. They were expensive. And because they were expensive, organizations learned to stop proposing them.
Research in the Creativity Research Journal found that organizations systematically lose innovative ideas because employees self-censor proposals perceived as too risky or too expensive [8]. When execution cost drops by an order of magnitude, self-censorship — not capability — becomes the binding constraint on innovation. Ideas rejected for cost deserve another look when that cost falls. A shortage of ideas and a shortage of budget are different problems, and only one of them just got solved.
Duolingo makes this concrete. The company spent twelve years building its first 100 language courses, each requiring teams of linguists, translators, and content designers working through months of development cycles. Growth was bottlenecked by what it cost to produce content — not by demand for it. In 2024, AI collapsed that cost. Rather than producing the same 100 courses with fewer people, Duolingo built 148 new courses in roughly one year. Daily active users grew 40% to 47.7 million; paid subscribers hit 10.3 million.
Twelve years for 100. One year for 148. The demand was never fixed. It had been suppressed by supply constraints that went unquestioned, because questioning them was pointless. The cost was what the cost was. Until it wasn’t. The same shape holds in software: GitHub reported 43 million pull requests merged each month in 2025, up 23% year over year alongside Copilot adoption [14] — not fewer engineers shipping the same code, but more code shipped because cheaper authorship made more projects worth starting.
I call it the “too-small-to-staff” problem. Every organization has a minimum viable project size set by human labor costs. Below that floor, you decline the work — it doesn’t pencil out. AI obliterates that floor. And the aggregate revenue of declined inquiries, too-small contracts, and not-worth-sending-someone opportunities is often larger than the work you currently accept. It was always there. You just couldn’t afford to see it.
The Ambition Audit
Here is what I tell clients sitting on the contraction side of this decision.
Surface the self-censored ideas. Ask your leadership: what did we kill in the last three years specifically because of execution cost — not because the market was wrong? Revisit the original reasons for rejection; the staffing estimate and the market assessment may tell different stories. The most valuable buried ideas are often older, from a time when the organization was more ambitious but more resource-constrained.
Map the adjacent markets. Capability adjacent: what else can you do with the skills you have, if execution cost were not a constraint? Customer adjacent: what else do your customers need that you never offered because you couldn’t staff it? Market adjacent: what segments look like your current customers but are too small or dispersed to serve at old costs? These are your Vitaminwater — categories that only appear when you stop defining your business by what you sell today. This is also where domain experts become builders: the people who know the work best can now prototype the new offering themselves, because the cost of building dropped to meet their expertise.
Identify the too-small-to-staff opportunities. Audit your minimum engagement size. What did you decline last year because it fell below your cost floor? That floor just dropped by 5x. Quantify the aggregate revenue — in many professional services firms, the declined work exceeds the accepted work in total volume. That is not a rounding error. That is a market you have been ignoring.
Run the expansion economics. Build two scenarios for the board: contraction versus expansion. Even with conservative assumptions — 30% of identified opportunities pan out, 12-18 month ramp — expansion typically wins on a five-year horizon because it compounds, while contraction is a one-time savings. The contraction case is concrete and easy to model; the expansion case requires faith in markets you haven’t served yet. Start with one pilot, show early results, and let the board extrapolate from proof rather than projections.
Rewire the planning assumptions. Your capacity model, market sizing, and hiring plans all embed execution-cost assumptions that are no longer true. If your plan assumes 40 hours per deliverable and AI brings it to 8, every downstream number is wrong. But update the incentives alongside the assumptions: if division heads are measured on margin percentage rather than absolute revenue growth, they will optimize for contraction regardless of what the plan says. The spreadsheet can show the opportunity. The organization will only pursue it if the people deciding are rewarded for expansion.
The leaders see this. A January 2026 EY-Parthenon survey found 69% of CEOs believed AI investments would maintain or increase headcount [3]; EY proved the model internally, consolidating operations through AI and then using the enhanced capability to create entirely new service offerings [10]. The Federal Reserve Bank of San Francisco expects AI to follow the historical pattern of transformative technologies that ultimately generate net new jobs and economic expansion [6]. Those are not the answers of executives planning to shrink the org. Boards setting strategy now have to decide whether those savings will fund expansion. Celebrating the lower cost alone leaves the larger opportunity unanswered.
The Defining Bet
The companies cutting headcount are betting the amount of work worth doing is fixed — that AI savings flow to the bottom line, and tomorrow’s market looks like today’s, only cheaper. The companies expanding are betting the pie was never fixed; it was constrained by what things cost to do. When those costs drop by an order of magnitude, the opportunity set doesn’t shrink. It explodes.
Capturing efficiency gains in the near term is fine — every well-run organization should. The error, expensive and compounding and possibly irreversible, is treating cost optimization as the entire strategy: pocketing the savings and declaring the job done while competitors reinvest those same savings into markets you haven’t imagined yet.
Jevons saw it with coal in 1865. Bessemer proved it with steel. Computing demonstrated it again with the internet, mobile, and cloud. Every time, the companies that treated efficiency as a license to expand outperformed the ones that treated it as a license to cut.
Your competitors have access to the same AI you do. If you are pocketing the savings while they are finding new markets, you are not winning an efficiency race — you are watching the pie grow from outside it. The ideas you killed are still there. The markets you declined are still there. The customers too small to serve are still there. The only thing that changed is the cost of reaching them.
Run the audit before your next planning cycle. The fixed pie was never fixed. It was just expensive to question.
References
[1] Rosalsky, G. (2025). “Why the AI world is suddenly obsessed with Jevons paradox.” NPR Planet Money. https://www.npr.org/sections/planet-money/2025/02/04/g-s1-46018/ai-deepseek-economics-jevons-paradox
[2] CNBC (2025). “AI was behind over 50,000 layoffs in 2025 — here are the top firms to cite it for job cuts.” https://www.cnbc.com/2025/12/21/ai-job-cuts-amazon-microsoft-and-more-cite-ai-for-2025-layoffs.html
[3] EY (2025). “Beyond Cost Cutting: AI as the Ultimate Growth Engine.” https://www.ey.com/en_us/insights/ai/beyond-cost-cutting-ai-as-the-ultimate-growth-engine
[4] EOXS (2024). “How 19th-Century Steelmaking Innovations Built Modern Industry.” https://eoxs.com/new_blog/how-19th-century-steelmaking-innovations-built-modern-industry/
[5] The Strategy Institute (2024). “Coca-Cola’s Global Dominance: Decoding the Beverage Giant’s Business Strategy.” https://www.thestrategyinstitute.org/insights/coca-colas-global-dominance-decoding-the-beverage-giants-business-strategy
[6] Daly, M.C. (2026). “The AI Moment? Possibilities, Productivity, and Policy.” Federal Reserve Bank of San Francisco, Economic Letter 2026-06. https://www.frbsf.org/research-and-insights/publications/economic-letter/2026/02/ai-moment-possibilities-productivity-policy/
[7] CEPR European firms AI survey (2026). Via secondary source. Preliminary data from study of 12,000 European firms on AI adoption and employment effects.
[8] Runco, M.A. & Acar, S. (2014). “Self-esteem and the self-censorship of creative ideas.” Creativity Research Journal. https://www.researchgate.net/publication/235254405_Self-esteem_and_the_self-censorship_of_creative_ideas
[9] Rosalsky, G. (2025). “Why the AI world is suddenly obsessed with Jevons paradox.” NPR Planet Money. (Jevons Paradox in water and transportation domains.)
[10] EY (2025). “Beyond Cost Cutting: AI as the Ultimate Growth Engine.” (EY internal AI Fabric transformation case.)
[11] Meta Platforms (2025). “Q1 2025 Earnings Call / Capital Expenditure Outlook.” Meta Investor Relations. https://investor.atmeta.com/financials/quarterly-earnings/default.aspx
[12] Microsoft (2025). “FY25 Q2 Earnings Materials / AI Revenue Commentary.” Microsoft Investor Relations. https://www.microsoft.com/en-us/Investor/earnings/fy-2025-q2
[13] Microsoft (2024). “2024 Work Trend Index Annual Report.” Microsoft WorkLab. https://www.microsoft.com/en-us/worklab/work-trend-index/2024
[14] GitHub (2025). “GitHub Copilot Research and Productivity Updates.” GitHub Blog. https://github.blog/tag/copilot/
[15] Brockway, P.E., Hamilton, I., Higgs, K., Barrett, S., et al. (2017). “A Review of Rebound Effects from Energy Efficiency and Conservation Policies.” Energy Policy. https://www.sciencedirect.com/science/article/pii/S0301421516307015
[16] Deming, D.J., Kahn, L.B., et al. (2023). “Artificial Intelligence, Productivity, and Employment: Evidence from Field and Labor-Market Studies.” NBER Working Paper w30783. https://www.nber.org/papers/w30783