The public market has been running this experiment for forty years and it has already priced the answer. Businesses whose cost of revenue is human time trade at 0.2–1.6x revenue, with the whole observed tail stopping at 2.3x — regardless of growth, brand, technology narrative or scale.
The cleanest single statement of it sits inside one company's filings. Robert Half's contract-staffing segment, which sells hours, ran a 39.0% gross margin in FY2025. Its permanent-placement segment, which sells a matched outcome, ran 99.8% (RHI FY2025 10-K). Same company, same salespeople, same buyers, same brand. The market pays 0.84x EV/revenue for the blend.
Every question on this page reduces to: which of those two segments is the business actually in?
The COGS test
If a human hour appears in cost of revenue and scales roughly 1:1 with revenue, you are a staffing, BPO or agency business no matter what the deck says. If human hours sit in S&M or R&D and don't scale 1:1, you may be software. The label does not determine what you are; the P&L does.
Palantir is the canonical exception that proves the rule. Its forward-deployed engineers are famously labour-intensive at the point of delivery, and FY2025 gross margin was still 82% (84% ex-SBC) (PLTR 10-K) — because the artefact sold is a licence that renews without proportional labour. The humans install a durable relationship; they are not the product. That is the template, and Forward-deployed engineering is where the atlas tests whether anyone else can execute it.
The table
All data 27–29 August 2026, stockanalysis.com.
| Company | Ticker | Revenue TTM | Gross margin | Op margin | EV/Rev |
|---|---|---|---|---|---|
| Palantir | PLTR | $6.16B | 84.8% | 42.8% | 71.2x |
| Airbnb | ABNB | $13.16B | 82.9% | 20.8% | 7.76x |
| DoorDash | DASH | $15.89B | 52.2% | 5.5% | 6.33x |
| Paycom | PAYC | $2.14B | 88.1% | 30.3% | 5.31x |
| ADP | ADP | $21.95B | 48.7% | 26.5% | 5.26x |
| Etsy | ETSY | $2.94B | 71.4% | 14.3% | 3.27x |
| Toast | TOST | $6.80B | 26.7% | 6.3% | 2.73x |
| Infosys | INFY | $20.30B | 29.7% | 20.4% | 2.28x |
| Accenture | ACN | $73.10B | 32.0% | 15.8% | 1.56x |
| Omnicom | OMC | $22.37B | 19.0% | 14.3% | 1.44x |
| Genpact | G | $5.25B | 36.5% | 15.0% | 1.39x |
| Cognizant | CTSH | $21.64B | 33.4% | 16.3% | 1.38x |
| Korn Ferry | KFY | $2.91B | 24.3% | 12.9% | 1.28x |
| C.H. Robinson | CHRW | $17.00B | 8.5% | 5.2% | 1.14x |
| Upwork | UPWK | $787M | 77.2% | 17.1% | 1.13x |
| Robert Half | RHI | $5.29B | 36.8% | 0.2% | 0.84x |
| Teleperformance | TEP.PA | €9.98B | 30.7% | 11.5% | 0.81x |
| Concentrix | CNXC | $10.00B | 34.5% | 6.2% | 0.70x |
| TaskUs | TASK | $1.23B | — | — | ~0.61 (P/S) |
| TTEC | TTEC | $2.04B | 23.3% | 4.2% | 0.45x |
| ManpowerGroup | MAN | $18.72B | 16.2% | 1.9% | 0.22x |
| Fiverr | FVRR | $418M | 82.0% | 9.7% | 0.09x |
The threshold is real and it sits around 40%
Every business below ~40% gross margin trades at ≤1.6x EV/revenue — with two exceptions in the table above, both of which have an explanation.
The ceiling case is Accenture: 32% gross margin, 15.8% operating margin, $73.1B of revenue, the best-run services firm on earth, and 1.56x. Forty years of flawless execution does not buy a services business a software multiple.
The two entries above 2x with a sub-40% gross margin are Toast (26.7%, 2.73x) and Infosys (29.7%, 2.28x). Toast's low margin is an artefact of gross-booked payments revenue sitting on top of a genuine software subscription business, which is the GMV is not revenue problem in reverse. Infosys has no such excuse: it is a pure offshore IT-services firm on a 20.4% operating margin, and it is the top of the BPO band described below rather than an escape from it.
The research note this table comes from asserts "exactly one exception above 2x" and names only Toast. Its own table lists Infosys at 2.28x on a 29.7% gross margin. The rule as stated is therefore a strong tendency with a visible upper tail, not a law — read the band as 0.2–2.3x for labour businesses, with 1.6x as the point above which you need a specific reason.
That is the ceiling the 2000s offshore BPO wave hit. Infosys at $20.3B, Accenture at $73.1B and Cognizant at $21.6B are the successful outcome of the largest labour-arbitrage trade in history, and it buys ~30–33% gross margins and 1.4–2.3x EV/revenue. The arbitrage was real and it was competed into a commodity, because the input was replicable and the buyer could always run a second RFP.
High gross margin is necessary and not sufficient
Upwork has a 77.2% gross margin, 17.1% operating margins and $204M of free cash flow, and trades at 1.13x — below Accenture. Fiverr has an 82.0% gross margin and trades at 0.09x: a $343M market cap against $308.5M of cash and securities, meaning the market values the operating business at approximately zero.
What killed them is not margin. It is visibility and volume. Fiverr guided FY2026 revenue at -17% to -14% with annual active buyers down 21.9% to 2.7M; Upwork's GSV is -4% YoY with active clients -4% — the same flat-volume picture that shows up as a rising rake in What a rake can actually be.
A software multiple requires gross margin ≳ 70% AND recurring or contractually visible revenue AND net revenue retention > 100% AND fragmented buyers AND positive volume growth. Miss any one of the five and you fall into the services band. Fiverr misses one and trades at 0.09x.
The businesses that escaped are instructive precisely because none of them escaped on narrative. ADP (5.26x, 48.7% GM) is a payroll BPO by origin that got out via regulatory-compliance lock-in, near-zero churn, float income and hundreds of thousands of price-taking SMB buyers. DoorDash (6.33x, 52.2% GM, 5.5% operating margin) is a genuinely labour-heavy managed marketplace earning its multiple on 27% GOV growth alone — growth is the one factor that can partly substitute for margin. And at the other end, TTEC at 0.03x price/sales ($69M of equity on $2.04B of revenue) is what leverage plus commoditised labour looks like at the end of the road.
What this implies about the private marks
There are exactly two public pure-plays doing something close to what the private AI-data set does.
| Anchor | Revenue | Multiple |
|---|---|---|
| Appen (ASX:APX) | US$232.7M FY2025, loss-making every year since 2022; TTM A$361.9M | ~0.9x revenue — A$330M market cap on A$361.9M TTM, AUD on AUD; do not divide the AUD cap by the USD revenue |
| Innodata (NASDAQ:INOD) | $317.2M TTM (+39%), ~40% gross margin, $32.2M net income FY2025 | 6.2x revenue ($1.95B market cap) |
Innodata is the bull case fully priced: growing 39%, profitable, and still only 6.2x on a 40% gross margin. Appen is the bear case fully priced: the same business, concentrated and shrinking, at 0.9x. Everything above 6x in the private register is a bet that the private companies are structurally different from these two.
Against that anchor:
| Basis | Range |
|---|---|
| Public AI-data comps | 0.9x (Appen) – 6.2x (Innodata) |
| Private, on gross billings | 3.2x (Handshake AI) – 33x (Scale AI on 2024) |
| Private, on net revenue | ~7.8x (Handshake AI) – ~38x (Mercor at $10B) |
| Private, on ARR, where the business is genuinely software-ish | 10x (AlphaSense) – 58x (Distyl AI) |
Distyl AI at $1.8B on ~$31M of ARR is the most aggressive mark in the register — a software multiple paid to a firm its own coverage calls an AI consulting startup, with 111 employees. But the $31M comes from GetLatka, a source class the research explicitly says to discard: the same estimator lists Surge AI — a company doing over $1B — at $330K of revenue and three employees. The multiple may be 58x, or the denominator may be wrong by a factor of anything. [WEAK]
The same caution applies to Toptal's widely-quoted ~$167M, which comes from the same estimator class and is the basis of nothing.
The load-bearing assumption in every mark above 6x is Sequoia's: that AI compresses the labour cost of delivery fast enough for gross margin to migrate from services levels (30–40%) to software levels (70–85%) before the multiple has to be justified (Sequoia, Services: The New Software). If that migration happens, 58x on Distyl is early. If it doesn't, it is 58x on a consultancy — and Accenture, at thirty times Distyl's headcount efficiency and a 32% gross margin, is trading at 1.56x.
Where the record is thin
Every multiple here is a single 2026 reading. Take rates and multiples are cyclical: Robert Half at a 0.2% operating margin is plainly at a cyclical trough, and Fiverr at 0.09x may be at a sentiment trough rather than a structural one. A 2015 / 2019 / 2021 / 2026 panel would separate structural from cyclical and has not been built. Adecco and Randstad could not be retrieved at all.
Two numbers used in arguments around this page do not fully reconcile. Palantir's gross margin is given as 82% (84% ex-SBC) from the FY2025 10-K in one note and 84.8% TTM in the market-data pull — close, different periods, and the 82% figure is flagged as unconfirmed against the filing itself. And GLG's "no client above 6% of revenue" — the concentration contrast used in One customer is a binary event and GLG — could not be corroborated in any research note; the unread October 2021 S-1 would settle it.