Gusto followed real payroll records for 1,593 small businesses that adopted AI and 669 that knew about it and chose not to, and found adopters grew headcount about 7% more over the following year (Gusto, 2026). That is the number that ran as proof AI creates jobs. The size breakdown underneath it is the number an operations leader should actually read: 10% at firms with one employee, 11% at two to four, 11% at five to nine โ and 0% at firms with 10 to 24 employees.
The dividend does not shrink gradually with scale. It disappears at the point where a business starts having processes.
That is not an argument that AI stops working above ten people. It is an argument about where AI freed capacity goes when it is not assigned a destination โ and the answer, consistently across three independent datasets, is that it goes into the existing process and stops being visible to anyone funding the pilot.
What 2,262 Payroll Histories Actually Show
The method is why this is worth more than another adoption survey. Gusto took respondents to its April 2025 AI survey, split them into 1,593 self-reported adopters and 669 aware non-adopters, and then measured what their payroll did โ not what they said it did. The comparison group is the careful part: businesses that knew about AI and declined it, which isolates the choice to adopt rather than comparing users against the unexposed.
The estimator is a Callaway and Sant'Anna event study, built for staggered adoption, with each business compared against not-yet- and never-adopting peers relative to the quarter before it adopted. Employment paths were flat and parallel before adoption โ the check that matters. The effect built gradually, from about 4.5% to about 6.8% across the four quarters after adoption, which rules out a one-off hiring spike. Wages showed no matching movement, consistent with firms adding people rather than riding a pay increase.
And the roles were not AI roles. Adopters hired therapists, teachers, technicians, cooks โ what Gusto calls do-ers, people who produce more of the thing the customer came for. Among firms under ten employees, headcount growth was 8% in goods-producing, 8% in personal services, 9% in professional services, and 19% in community services, driven by small health care and social assistance practices hiring dental hygienists and clinicians.
This is a clean result with an unglamorous mechanism: in a five-person business, the owner's hours are the binding constraint, and those hours are spent on scheduling, admin, marketing and research. AI absorbs that, and the recovered hours convert into the next hire.
The Zero Is the Finding
Gusto's own explanation of the 10-24 band is the most operationally useful sentence in the study. Firms that size, it notes, "already have much of these processes in place," and their productivity gains "likely show up in other places than headcount, even if employees are freed to focus more on their core jobs" (Gusto, 2026).
Read that twice, because it is the whole pilot-to-budget problem stated by the people holding the data.
At five employees there is no process to absorb the freed hours, so they surface as an observable decision: hire someone. At fifteen, there is process โ a scheduling system, an approval chain, a reporting cadence, a queue of work that expands to fit available attention. The hours still get freed. They are simply reabsorbed by the structure that was already there, and no adoption dashboard, usage metric or satisfaction survey can tell you where they went.
Scale that forward to 50-500 FTE and the structure is denser by an order of magnitude. The assumption sitting quietly under most 2027 AI business cases โ that capacity will find its own best use โ is precisely the assumption that measured zero at fifteen employees.
Denmark Shows Where the Freed Capacity Went
If capacity is reabsorbed rather than destroyed, something should be absorbing it. The best evidence on what comes from administrative data rather than self-report.
Humlum and Vestergaard linked large-scale AI adoption surveys to Danish administrative labour records and estimated precise null effects on earnings and recorded hours at both the worker and workplace level, ruling out effects larger than 2% two years after ChatGPT's launch โ this in a setting where most employers in exposed occupations had adopted chatbot initiatives and workers reported productivity benefits (Humlum & Vestergaard, 2026).
What moved was the structure of work. Employers absorbed AI through task reorganisation, and the authors name the new tasks explicitly: content generation, AI oversight, and AI integration. Adopters also moved into higher-paying occupations where chatbots are more relevant, though too few to shift average earnings.
That is the destination, and it is not the one in the business case. The time the tool gives back is substantially spent supervising, checking and wiring up the tool. Nobody wrote that line into the budget, nobody owns it on an org chart, and because it is distributed across everyone's day in ten-minute increments, it never appears as a cost โ only as a productivity gain that somehow did not land.
The Gains Above the Line Are Real โ and They Are Bought
The honest counterweight to Gusto's zero comes from a much larger firm panel. Using matched EIBIS-ORBIS data on more than 12,000 non-financial firms in the EU and US, instrumenting EU adoption with the adoption rates of US peers, AI adoption raises the level of labour productivity by 4%, through capital deepening rather than job displacement, with no adverse firm-level employment effect (BIS, 2026).
Two details matter more than the headline. First, the gains "concentrate in medium and large firms" โ exactly the band where Gusto's headcount effect vanished. Second, the paper is explicit that complementary investments in software, data and workforce training are what unlock them.
Put the three datasets in one line and the picture is coherent. Below ten employees, freed capacity converts itself into headcount because there is nothing in the way. In medium and large firms it converts into measured productivity โ but only alongside the complements that redirect it. In between, and in any firm that buys the tool without the complements, it converts into unowned task overhead.
The dividend did not die at ten employees. It stopped being automatic.
The Honest Limits
Gusto is associational, and says so. Businesses choose whether and when to adopt; adoption timing is self-reported; the sample is Gusto customers who answered a survey, not a random draw of US small businesses. The authors state plainly that exact magnitudes may not hold economy-wide.
The 5-9 band carries a caveat from the authors. Those firms were slightly shrinking before adoption, so part of that 11% may be bounce-back rather than new hiring.
The zero was measured at 10-24 employees, not at 200. Nothing in this dataset observes a 300-person operation. What travels upward is the mechanism โ process absorbs unassigned capacity โ not the point estimate. Anyone who quotes "0% above ten employees" as a finding about the mid-market has overread it, including anyone citing this article.
Denmark is early, chatbot-specific, and about averages. A precise null on average earnings and hours is not a null on any individual role, and the window is the first two years of a general-purpose technology.
The BIS estimate identifies exposure, not your rollout. Instrumenting with US peer adoption rates buys clean causal variation at the cost of measuring the average effect of being an adopter, not the effect of how well you implemented.
What a Head of Operations Should Do Before the Next Pilot Is Funded
Four moves, and none of them is a tooling decision.
1. Write the destination of the hours before approving the spend. One sentence, in the approval: the capacity this frees goes to this output, owned by this person, measured this way, by this date. If the only available answer is "the team will be more productive," you have just funded the 0% condition โ a freed hour with no named claim on it is reabsorbed by the process that was already there.
2. Budget the oversight and integration work as work. The Danish evidence says the new tasks are supervision, verification and wiring-up. Price them into the business case as a named cost against a named role, or they will be paid out of the same capacity the pilot was supposed to release.
3. Change the dependent variable you report. At your size, headcount is the wrong readout and usage is a vanity metric. Pick something the process cannot absorb silently: cycle time on a specific workflow, backlog cleared, output per FTE in one function. Instrument it before the pilot starts, because a baseline you reconstruct afterwards is a story.
4. Give the oversight a name rather than an overflow. The reason verification work disappears is that it has no seat. A governed deployment makes it structural instead: Scovai's Digital Employee model is confirm-first by default โ every externally visible action becomes an approval request with a named human on it, autonomy is earned per kind of action from a recorded ledger, and silence is never consent. Whatever stack you choose, the test is the same: can you point at the person who owns the approval queue, and can you see what it cost them?
The three datasets disagree about magnitudes and agree about the mechanism. AI freed capacity that nobody has claimed is not saved. It is spent quietly, inside the structure you already have, on work nobody decided to do.
So pick one AI pilot that is already running in your organisation and try to finish the sentence in move one. If you cannot name the destination, the owner, the measure and the date this quarter, you do not have a productivity programme. You have a reallocation you are not watching.
