Workers in the smallest workspaces use AI for somebody else's job 18.9 percent of the time. In workspaces above a hundred seats, the figure is 16.3 percent. Those two numbers come from OpenAI's first Work at the Frontier report, published this week from a sample of more than 800,000 work-related messages, and the report treats the gap as evidence that small organisations lean harder on AI because they have fewer specialists to call on.
Put differently: the large firm is running at 86 percent of the small firm's rate of working outside its own job description.
This is interesting, because those two kinds of organisations were never supposed to resemble each other on this measure at all. A five-person company has always operated on a single principle, which is that there is nobody to hand it to. The person who notices the problem is the person who fixes it, whether or not it belongs to them, because the alternative is that it does not get done. Nobody called this a coordination model. It was just what small looked like.
The large firm was defined against exactly that. Specialist attention was expensive and it queued, so the organisation existed to route work toward whoever owned the relevant competence. The handoff was the unit: the marketer briefed the developer, the salesperson sent the dataset to the analyst, and the org chart was a map of where the routing went. Strip away scale, revenue, and floor space, and the operational difference between a large firm and a small one was that the large one could hand things off.
OpenAIs report gives no time dimension, so this is convergence observed rather than diffusion measured, and the seat count is not company size either, which the report says outright. But that second caveat sharpens the argument: what a seat count measures is how many colleagues sit inside the same working perimeter, which means a five-seat team inside a large bank is already the small firm for this purpose. Proximity to somebody you could hand work to is the variable. Corporate headcount was never it.
The strongest evidence for the convergence is not the gradient anyway. It is a table OpenAI published on the web page but left out of the PDF, showing how each occupation's own work divides across the eight possible task sources. Rows sum to a hundred, and the diagonal is the share of a worker's occupation-specific messages that concern their own occupation's traditional work.
| Worker occupation | Cust. exp. | Design | Eng. | Finance | HR | Legal | Marketing | Sales |
|---|---|---|---|---|---|---|---|---|
| Customer experience | 11 | 5 | 20 | 14 | 6 | 6 | 26 | 11 |
| Design | 6 | 12 | 28 | 10 | 4 | 4 | 28 | 8 |
| Engineering | 4 | 4 | 53 | 9 | 4 | 4 | 20 | 2 |
| Finance | 6 | 4 | 22 | 23 | 4 | 8 | 25 | 7 |
| Human resources | 9 | 5 | 18 | 16 | 10 | 10 | 23 | 9 |
| Legal | 5 | 4 | 17 | 12 | 7 | 31 | 18 | 6 |
| Marketing | 10 | 6 | 17 | 11 | 4 | 4 | 36 | 12 |
| Sales | 11 | 5 | 18 | 15 | 4 | 6 | 29 | 12 |
In a labour market organised around specialisation, that diagonal should dominate every row. It dominates three. Engineering holds 53 percent of its own composition, marketing 36, legal 31. For the other five occupations the single largest source of their occupation-specific work is somebody else's job, and in four of those five it is marketing. Human resources holds ten percent of itself.
This is what a generalist looks like from the inside, and it is what small firms have always looked like. The report captured it happening at scale.
Whether the tool is causing it or merely coinciding with it has one reasonable test, from a study Perplexity published in June with Harvard Business School. Across 8,000 users spanning eight occupation clusters, the same people worked outside their primary occupation 59 percent of the time when using the agent product and 50 percent when using conversational search. Same users, both products, nine points of difference. Perplexity co-authored a study of its own product using its own production data, it has not been peer reviewed, and the authors note their early window skews toward AI natives, so treat the magnitude as indicative. The direction is harder to dismiss.
The more useful detail is where the extra crossing went. With search, cross-occupation queries concentrated into digital technology, which is the familiar pattern of everyone occasionally needing a technical answer. With the agent, they spread across marketing, management, financial services and other destinations the authors describe as executional. Search sent people to look something up outside their job. The agent sent them to go and do it.
Which work makes that trip is not random, and the matrix explains the selection better than any account based on difficulty. Troubleshooting a computer application ranks among the three most common outside tasks in every non-engineering group in the sample, and it travels because failure announces itself: the error stops or it does not. What stays behind engineering's 53 percent is the work where failure is delayed, and an outsider cannot attempt that, because they would not know for months that they had got it wrong.
Marketing has almost none of that structure, which is why it is the largest external source in five of eight rows. A mediocre landing page looks like a good one on inspection. Design is the case that proves the point rather than merely fitting it: design output is also judged by taste rather than by tests, so a difficulty account predicts it should travel as freely as marketing, and instead its column is the thinnest in the table at 4 to 6 percent everywhere outside its own row. A bad layout announces itself on sight. Loud failure keeps outsiders out. Silent failure lets everyone in.
So the small firm's model has arrived in the large firm, and it arrives task by task, in the order set by how quickly a non-specialist finds out they were wrong.
The firms themselves are responding, and not in the direction most commentary assumes. A Purdue study published in May built an AI exposure measure across more than nine million US job postings from 2021 through mid-2025, letting exposure vary within an occupation over time rather than fixing one score per title. From the third quarter of 2023 onward, 52 percent of the decline in aggregate exposure came from firms shifting hiring away from exposed jobs, and another 39.5 percent came from rewriting the task content of jobs they continued to post.
Individuals are pulling exposed work in. Employers are pushing it out of the job descriptions. Both are the same organisation adapting, and what the second half tells you is that firms are already redefining roles around whatever survives the question of whether the output can be trusted without a specialist looking at it.
There is also reason to think this is closer to its ceiling than its beginning. Anthropic's economic index reported that its estimate of jobs with at least a quarter of tasks touched by AI barely moved between successive data pulls, with far fewer novel task types appearing than in the previous one. The breadth of work being reached looks like it is saturating.
None of which is cost-free, and the cost is the thing the small firm never had to lose.
The handoff was doing two jobs at once. It routed work to the person who could do it, and it routed work to the person who could tell whether it had been done properly, and those arrived bundled because for most of industrial history they were the same hire. The large firm bought both and only ever noticed it was paying for the first. Remove the routing and the second goes with it, silently, because nothing in the org chart was ever labelled review.
Small firms have run without that review forever and developed the instincts to survive it, which mostly means knowing which of your own outputs you should not trust. Large firms have no such instincts, because they never needed them, and they are now acquiring the small firm's working pattern without the small firm's scepticism about it.
For anyone building agent systems the same collapse is happening one layer down and considerably faster. Every delegation to an agent is a non-specialist attempting somebody else's work, and the property that decided which tasks travelled between humans decides which delegations are safe. A migration that runs against a test suite the agent did not write is engineering's loud failure, and it stops. A summary, a plan, a recommendation: silent, fluent, and indistinguishable from the good version until much later.
The report's own conclusion is that government occupation statistics will drift away from how work is really organised. That is true and it is the smaller half. Every internal dashboard counting runs, completions and tasks attempted is measuring the half that got cheap, and the Purdue data leaves one detail for anyone who thinks they know which way this runs: senior postings carry higher AI exposure than junior ones, 0.582 against 0.426. The people assumed to be safe are doing the exposed work.
Unlock the Future of Business with AI
Dive into our immersive workshops and equip your team with the tools and knowledge to lead in the AI era.