Usage is easy to count and tells you little. Revenue per employee is one number analysts watch to judge AI, and it is easy to misread. Grindr is the test case.
THE APP THAT TRIMMED ITS BEST STATISTIC
In 2023, Grindr gave its staff a choice. Come back to the office two days a week, or leave.
A lot of them left. The union counted about 80 of 178 people, or 45%, by the end of August 2023. The CEO later estimated that 60% of staff left. It looked like a self inflicted wound.
What remained was a small team, a 17 year old codebase and a reputation. Grindr’s leadership would later say the app’s bugginess “was like a meme in gay social discussions all the time.”
$195.0 million of revenue in 2022. More than 200 full time employees.
$439.9 million of revenue in 2025. 165 employees.
Do the sum. In 2022 each employee stood behind $0.97 million of revenue. By the end of 2025 it was $2.67 million. A dating app with 165 employees had nearly tripled a number analysts watch closely.
The company had a story about why. In its second quarter letter of 2026, Grindr said engineering output had risen about 2.5 times in nine months, with minimal growth in the technical team.
The reason, it said, was AI agents. Cursor. Claude Code. Devin. The extra output, it said, equals about 200 more engineers. About $60 million a year.
The earnings call added a twist. The CEO said the true figure was 3.5 times. The company had cut it to 2.5 because 3.5 seemed unreasonable.
A company trimmed its best statistic because it feared nobody would believe it.
Is any of it true? Grindr is unusually open about its numbers, which makes it a good test. Look at how the 2.5 times was measured. Mainly by code shipped, the CEO said on the Q2 call.
Apple tried measuring engineers that way in 1982. I come back to how that ended.
This article is not about whether a company uses AI. Nearly all of them do. It is about whether real business metrics moved because of it.
MOST COMPANIES CANNOT SHOW THE EFFECT
Companies publish what is easy to count: seats bought, weekly logins, tokens burned, the share of code written by AI. These only go up, and nobody outside can check them.
The outcome data is thin. McKinsey surveyed 1,719 people in mid 2026. Only 37% said AI had added anything to EBIT, the same share as before. BCG’s September 2025 survey of 1,250 executives found that 5% of companies get AI value at scale and 60% report minimal revenue and cost gains.
All of that is self reported, and it points one way. Most companies cannot yet show the effect in their numbers.
Investors fall back on outside numbers. Most need private data: burn multiple, retention cohorts, margin by product line. One needs only a filing: revenue per employee.
REVENUE PER EMPLOYEE: ONE NUMBER ANALYSTS WATCH
AI’s core promise is the same people, or fewer, producing more. If that holds, revenue per employee, which is revenue divided by headcount, rises.
The ratio is not new. Tomasz Tunguz charted the average of 12 leading public software companies, with revenue per employee rising from about $200,000 in 2013 to about $470,000 in 2023. That climb began years before generative AI reached the office.
AI changed who was asking. A number from venture memos started appearing in press releases. In May 2025 Klarna said its revenue per employee was on track to reach $1 million, and TechCrunch put the figure a year earlier at $575,000.
Fortune projected that Block’s gross profit per employee, a stricter measure than revenue, would reach about $2 million in 2026 if it met its outlook.
Benchmark pieces followed. SaaStr, summarizing a16z data whose population it does not specify, said the 75th percentile had nearly doubled to $350,000 of ARR per employee and the 90th was pushing $700,000.
The ratio is easy to misread. Layoffs move it. Contractors hide in it. Pricing flatters it. That is why research struggles to link it to AI. One study covered 510 S&P 500 companies over 4,561 firm years from 2016 to 2025.
It scored each firm’s AI adoption from the text of its 10-K. It found no significant association between AI adoption and revenue per employee. That is description, not proof of cause.
An OpenAI report, as Fortune covered it in August 2026, found revenue per employee not meaningfully associated with output tokens per employee. Fortune notes the revenue figures predate the employees’ ChatGPT use, so the data cannot show cause.
A GTMnow analysis argues that the biggest gains belong to the companies that cut the most.
Here is where companies sit.
| Company or group | Revenue per employee (approx.) | Note |
|---|---|---|
| Ten large SaaS companies, average | $610K | Jeremiah Owyang’s May 2025 post, using Lean AI Native Leaderboard data. Unaudited |
| Ten lean AI startups, average | $3.48M | Same post. Curated winners, so biased upward. Midjourney alone adds about $1M |
| Microsoft | $1.2M | Big tech floor. FY2025 10-K: $281.7B revenue over about 228,000 employees. My arithmetic |
| Alphabet | $2.1M | FY2025, 10-K |
| Grindr | $2.7M | FY2025 on 165 employees, 10-K |
| Nvidia | $5.1M | FY2026, year ended January 2026, 10-K |
Headcount dates differ. Revenue periods differ. The two benchmark rows come from a hand picked list of ten lean AI startups against ten large SaaS firms. Without Midjourney, the AI average falls to about $2.5M.
Three versions of the metric are worth knowing.
Revenue per employee. The headline. Easy to compute, easy to distort.
- Gross profit per employee. Harder to flatter. It matters most for AI native companies, whose model bills sit in cost of revenue.
Revenue per labor dollar. Revenue divided by total people cost. It catches the case where a company cuts heads but pays the rest much more. Klarna’s average compensation went from $126K to $203K while its headcount fell by about half.
One rule. The direction matters more than the level. A company at $600K that climbs for eight quarters is telling you more than a company parked at $3M.
The benchmarks have a bigger problem. For the most talked about AI companies, reliable data barely exists.
Cursor, the poster child of tiny teams, has never published its numbers. Press reports cite roughly 150 to 300 people and annualized revenue above $2 billion to $4 billion, so revenue per employee could be anywhere from $7 million to $27 million.
Midjourney is no better, and aggregator sites are guessing. Use filings where you can and label the rest as estimates.
The private company numbers are too thin to trust. We need a company whose filings let us check the work. Back to Grindr.
GRINDR: THE STRANGEST TEST CASE OF 2026
A claim is only a claim. Here is what the filings show.
Revenue of $195.0M, $259.7M, $344.6M and $439.9M, divided by year end employees, from Grindr’s 10-K filings. The ratio is my arithmetic. 2022 counts full time employees only; in 2025, 160 of the 165 were full time. The 2026 bar (hatched) annualizes first half revenue over 172 full time US staff, a different headcount definition, so treat it as a rough guide.
Look at 2023. Revenue per employee more than doubled in one year. AI is not the full story. Revenue rose 33% while the headcount collapsed.
The AI story starts in 2025. From 2024 to 2025, revenue per employee rose 14%. In 2026, annualized first half revenue over 172 US full time staff gives about $3.1 million.
That uses a different headcount definition, so treat it as a rough guide. The AI era gain is real. It is about 30% over two years.
The filings are only half the story. Grindr’s claims about AI are much bigger.
The company describes the rebuild in detail. Press accounts say up to 80% of its code is now written by AI. The chief product officer said 70% is checked in via AI.
Nobody defines written. The engineering blog itself puts Cursor at about 30% of the team’s code. The engineering blog reported a January 2026 survey of 50 of 65 product engineers. 94% ran 1 to 5 agents at once. 92% said productivity was up 1.5x or more. All of it self reported.
How was output measured? Asked by a Morgan Stanley analyst, the CEO said, “we looked at how much was shipped in a period of time.” That is a volume measure, reported by the company and not audited.
Adjusted EBITDA margin slipped from 45.0% in Q1 to 41.7% in Q2. Whatever AI saved, it did not show up as a higher margin that quarter.
One more number nobody else names. The CEO told CNBC, as Business Insider reported it, that Grindr was on track to spend about $6 million on AI tokens this year.
Set that against the $60 million of avoided hiring the letter cites. That is a counterfactual, not a line in the accounts. On the company’s own numbers, the claimed return is 10 to 1.
$6 million in tokens. $60 million of hypothetical avoided hiring.
On jobs, the CFO said on the call that Grindr is not looking to cut headcount. The letter says it expects to finish 2026 with meaningfully fewer employees than planned.
The CEO said he might have hired 10 or 20 more engineers. AI shows up as hiring not done, not as jobs cut.
The new bottleneck, leadership says, is product management. Engineers can now build faster than the company can decide what to build.
Is it real? Partly. Adoption and cost are documented. Speed is the company’s own estimate. Quality is missing, with no defect rates and no independent data.
The evidence is good, though it is not clean. My read: revenue per employee is up about 30% in two years. That is real, and a long way from 2.5x. The two numbers measure different things, and the filings cannot bridge them.
HOW REVENUE PER EMPLOYEE GETS MISREAD
Press on it and three problems appear.
Contractors. Headcount counts employees. Grindr’s 10-K describes a separate contractor team in Colombia: 20 full time engineers in 2024 and 29 in 2025.
Add them in and 2025 revenue per head falls from $2.67M to about $2.27M. Employee compensation and contractor expense, excluding stock based pay, rose 56% in Q2 2026, per the 10-Q.
That sits awkwardly beside the letter’s line about minimal growth in the technical team, and the call’s phrase “roughly the same size team.” The company has not said whether contractors count as the technical team. None of this makes the story false. It makes the number softer.
Revenue is not all productivity. Grindr’s advertising revenue grew 44% in Q2. Higher prices began rolling out in Q3 2025. Average paying users grew 16%.
Some of the 33% growth came from charging more and selling more ads, not from shipping faster. Growth has already slowed, from 38% in Q1 to 33% in Q2. AI may have helped. The metric cannot tell you how much.
The layoff effect. Cut heads, keep revenue, and the ratio improves immediately. That tells you about cost, not about AI.
The layoff effect looks like AI from the outside. Block is the clearest case.
In February 2026, Block cut from just over 10,000 people to just under 6,000. Dorsey tied the cuts to AI.
Dorsey explained the decision on Sequoia’s podcast: the minimum people needed to keep the service up, to stay compliant, and to meet commitments to the Street.
Jack Dorsey on Sequoia Capital’s Long Strange Trip podcast, where he describes the three question headcount exercise behind Block’s February 2026 cuts.
Block had grown from 3,835 people at the end of 2019 to 12,985 at the end of 2023. That is 3.4x in four years. Dorsey conceded on X that Block over hired during covid.
The cuts did two things at once. They corrected an over hire, and they raised output per head. From outside, you cannot tell how much was AI.
Gross profit per employee is the fairer lens here, because bitcoin pass through revenue inflates revenue per employee.
Figures as reported by Fortune, with the 2019 number attributed to Block’s COO. My check against the filings is consistent with about $1.0M for 2025. The 2026 bar (hatched) is Fortune’s projection if Block meets its outlook, not a result. Block cut from just over 10,000 to just under 6,000 people in February 2026.
Block also claims code changes per engineer are up 150% since the start of the year. That is a management claim, not an audited one.
Block is not the only case. Shopify looks like a clean AI case.
In April 2025, CEO Tobi Lütke told teams to show why AI could not do the work before asking for more headcount. It reads like an AI story.
The cuts came first. Shopify cut about 1,000 people in 2022 and about 2,000 in May 2023. Headcount fell from about 11,600 at the end of 2022 to about 7,600 at the end of 2025. The memo came after the big drops.
Revenue per employee went from $483K to $1.52M. Gross profit per employee went from $237K to $731K.
Revenue still grew about 64% from 2023 to 2025 on flat to falling headcount. The AI effect is real. It sits on top of a layoff story.
The check is simple. Compare the date headcount fell with the date the AI claims began.
Revenue per employee tells you whether the numbers moved, not why or at what cost. Part 2 covers the signals I read beside it, and how to check a company in an afternoon.
THE MINUS 2,000 TEST
The habit behind that 2.5x is not new. Apple tried to measure its engineers by volume in 1982.
In February 1982, the managers at Apple had a new idea. Every engineer on the Lisa team would fill out a form each week. It asked one thing. How many lines of code did you write?
Bill Atkinson was the programmer behind QuickDraw, the graphics engine under the Lisa and later the Mac. That week he had rewritten a core piece of it. The new version ran about six times faster. It was also shorter. By roughly 2,000 lines.
In the box marked lines written, he put minus 2,000.
The managers stopped asking Atkinson to fill in the form.
Atkinson’s good week scored minus 2,000. A board that ranks people by volume would have put him last.
Here is the test. When a company tells you how much AI it uses, ask for its minus 2,000. What went down while value went up? Code removed. Bugs closed. Cost per release. Heads per dollar of revenue. Hours lost to rework.
A company that can show you a number going down is usually measuring something real. A company that only shows you numbers going up is usually measuring itself.
Ask the minus 2,000 question of your own team. What went down while pipeline went up? Cost per qualified meeting. Hours spent building lists. Days from signup to first human touch.
If none of those moved, the AI is producing volume, not value. To design the test properly, start with GTM experiments by stage.
WHAT TO TELL YOUR CEO
When someone says AI made the company more efficient, ask three things. When did headcount change? Which number went down? Who audited it? If the answers are before the AI, none, and the company itself, you have a story, not a result.
If you run one of these programs, or you have seen a metric that actually held up, tell me. angkan.mukherjee@gmail.com