66 companies. 3 you can divide.

Revenue per employee is the flagship number of the AI-native movement. It is in the decks, in the benchmark posts, in the headcount plans. It is also, in most cases, uncomputable, and the reason is boring and structural rather than scandalous.

The metric needs two inputs. One of them is well sourced and one of them is not.

We keep verification records on AI-native companies in this repo, one file per company, each claim individually labeled verified, plausible, thin or contradicted against the source we could actually retrieve. On 2026-08-19 that corpus held 66 records and 516 labeled claims: 273 verified (53%), 124 thin (24%), 86 plausible (17%), 32 contradicted (6%), and one forecast ^[1].

Sorted by what the claim is about, the failures are not evenly distributed:

The gap between the top row and the bottom row is the whole article. A funding round arrives with a press release, a filing, a lead investor with an incentive to confirm, and three outlets racing each other to publish. A headcount arrives with a data aggregator and a LinkedIn page. Only 9 of the 66 records carry a verified claim that states an actual number of employees at all ^[1].

So the metric everyone plans against has one input that verifies at 81% and one that verifies at 19%.

An automated pass over the corpus, looking for records with both a verified claim containing a numeric headcount and a verified claim containing a numeric revenue figure, returned 5.

We then re-fetched every underlying source, one at a time, and read what the page actually says. This step is not ceremony. A verification record is a secondary source, and a URL that resolves is not evidence that the page contains the claim attached to it. Two of the five fell out.

Decagon fell out on both inputs. Its verified "300+ employees" describes who was eligible to sell vested shares in a March 2026 tender offer, not a headcount, and its revenue is reported as having "surpassed eight figures," which is a category, not a number you can divide ^[7].

Glean is the more instructive drop, because on the surface it looks like the strongest case in the set. Its $300M ARR is in TechCrunch on 28 May 2026, a three-fold increase from $100M fifteen months earlier, and TechCrunch attaches its own caveat: because Glean prices partly on consumption, the milestone "cannot be fully described as traditional ARR" ^[5]. That article carries no headcount at all. The headcount in circulation is founder Arvind Jain saying the company "employs more than 1,000 people today" during a podcast appearance, relayed by a news aggregator ^[6]. More than 1,000 gives you a ceiling, under $300K per employee, and it is a floor on the denominator rather than a measurement of it. That is a bound. It is not a ratio.

Three held.

Gamma. $100M ARR, reached profitably, at a $2.1B valuation, announced 10 November 2025. On headcount, TechCrunch quotes the founder directly: "This was achieved with only about 50 employees" ^[2]. That is roughly $2.0M per employee ^[8].

Lovable. $400M ARR crossed in February 2026, reported 11 March 2026, with "just 146 full-time employees" ^[3]. Roughly $2.74M per employee ^[8]. Worth carrying the provenance detail: TechCrunch had the ARR confirmed by the company directly, while the 146 figure traces to the company's chief revenue officer speaking to Business Insider, and the article notes no independent verification of either ^[3].

Legora. $100M ARR, told to Business Insider and reported 1 April 2026, with headcount grown "from 40 to more than 400" over the same climb, attributed to the company ^[4]. Under $250K per employee ^[8].

Put Gamma and Legora side by side and the benchmark dissolves without any help from us.

Both crossed $100M ARR. Gamma did it in November 2025 with about 50 people. Legora did it five months later with more than 400. Same revenue line, roughly eight times the headcount, both figures given by the companies themselves to real publications, neither of them disputed by anyone.

Across the three computable pairs the range runs from under $250K to roughly $2.74M per employee. That is about 11x ^[8]. A spread that wide is not a benchmark with noise around it. It is a spread, and the median of a spread that wide tells you nothing about your own company.

There is a reasonable objection: these are different businesses. Gamma sells a self-serve consumer and prosumer product with 70 million users ^[2]. Legora sells to law firms, which means solutions engineers, security reviews and a physical presence in Stockholm, London, New York and elsewhere. Of course the ratios differ. That objection is correct, and it is the point. Once you accept that the ratio is a function of the go-to-market motion rather than of being AI-native, "AI-native revenue per employee" stops being a benchmark and becomes a description of what somebody happened to sell.

The count is ours and it is reproducible against this repo. Claims were extracted by splitting each record on its ### <n>. <label> headers and reading the Claim: line under each. Subject buckets were assigned by regular expression on the claim text: headcount on employees, headcount, staff, people, FTE and team-of constructions; revenue on ARR, revenue, annualized and run rate; funding on raised, Series, seed, valuation, funding and post-money. A claim mentioning two subjects lands in both buckets, so the bucket totals sum to more than 516.

Two known limits, stated rather than buried. First, the buckets are lexical, so a funding claim that mentions an employee tender offer matches the headcount pattern. That is exactly how Decagon and Clay entered the automated shortlist of 5, and exactly why the manual re-fetch cut it to 3. Second, the corpus is our reading list, not the universe of AI-native companies. It is 66 companies we chose to look at, so the verification rates describe how well this movement's claims are sourced, not a random sample of the sector.

Both limits push the same direction, and it is worth saying which way. A larger corpus would add more companies with a verified revenue figure and no verified headcount, because that is the shape of the underlying reporting. The ratio of computable to total would get worse, not better.

This is a blueprint problem, not a gotcha, so here is the blueprint part.

Do not set a headcount target from an industry ratio. There is no industry ratio. There are three companies whose ratio you can check and they disagree by an order of magnitude, and the variable that separates them is how they sell, not how much AI they run.

When you read a revenue-per-employee figure, ask which number was estimated. Almost always it is the denominator. The revenue figure usually traces to a company disclosure in a named outlet. The headcount often traces to an aggregator, and aggregator headcounts on the same company routinely disagree by 3x or more, which is a pattern this record has documented repeatedly across the past month.

Instrument your own denominator before you benchmark it. Your headcount is the one number in this entire calculation you can measure exactly, today, at zero cost. The comparables cannot say the same, and that asymmetry is worth more than any peer average.

And publish both numbers together if you publish either. Gamma's founder gave the revenue and the headcount in the same sentence, on the same day, at the same milestone ^[2]. That single sentence is why Gamma is one of three companies in this corpus whose efficiency claim can be checked by anyone who wants to. It costs nothing and it is the entire difference between a number and a claim.

None of this accuses any company here of misreporting. Every figure above is a real disclosure to a real publication, and where the provenance is second-hand we have said so. The narrower finding is the one that outlives the quarter: the flagship efficiency metric of the AI-native movement rests on its least verifiable input, and the people publishing industry averages for it have mostly not checked whether the division was possible.

We checked. It was possible three times out of sixty six.

Nothing is asserted that isn't shown.