The people getting the most from AI agents are not the engineers

The biggest gains from AI agents in a professional firm are most likely to come from the people who never touch code. Legal, finance and recruiting teams, not engineers, are now the fastest-growing users, according to a year of evidence OpenAI published this week. The practical move for a leader is to stop treating agents as an engineering tool and start asking which fee-earning and back-office tasks your own non-technical people could hand over, with a named person still accountable for checking the result.

That finding comes from OpenAI's research on how its Codex agent is used, released on 25 June. It is worth reading with a clear head, because OpenAI is reporting on its own product and, for the internal numbers, its own staff. Treat it as evidence with an interest attached, not gospel. The direction it points in is the durable part.

What the OpenAI numbers actually show

Three things stand out. First, non-developer use grew far faster than developer use: since August 2025, weekly non-developer users rose about 137 times among individuals and 189 times among organisations. Second, inside OpenAI itself, legal, finance and recruiting crossed over to using the agent as their main AI tool around April 2026, later than engineering but faster once they started. Third, the work people hand to an agent is getting bigger. By May 2026, roughly 81 percent of sampled individual users had made at least one request the system estimated would take a person more than 30 minutes, and about 70 percent had handed over something estimated at more than an hour.

A note on those last figures, because honesty matters here. OpenAI estimated the human time each task would take using a model, on a small random sample, and says plainly the numbers are directional rather than exact. So do not bank the precise multiples. What survives the caveats is the shape of the change: agents are spreading fastest among people who are not technical, and they are being used for longer pieces of work rather than quick lookups.

When the unit of work becomes a whole task

For two years the everyday use of AI has been the quick question. Draft this email. Summarise this report. Tidy this clause. Useful, but small, and always under the user's hand from start to finish.

An agent changes the unit. You hand over a task with a goal and some context, it works for a stretch, and it comes back with something to check. The closer parallel is briefing a junior colleague than typing a search.

For a professional firm, that is where it gets concrete. A recruiter can hand an agent a messy set of notes and a brief and ask for a first-pass market map and a longlist with sources, rather than typing one candidate query at a time. A finance analyst can point an agent at a folder of statements and ask for a first variance summary with the workings shown. An advisory associate can ask for a research pack on a sector, pulled together and cited, as a starting draft. None of these people writes code. All of them are now plausibly among the heaviest users in the building, because the work they do is exactly the kind that takes an hour by hand and can be delegated.

The OpenAI data hints at one more thing worth watching: more than a quarter of the agent work done by people in business functions was itself technical, things like pulling data together or building a small tool. Non-technical staff are quietly doing adjacent work that used to need a specialist. That is an opportunity and a supervision question at once.

The skill that matters now is clear delegation

If the unit of work is now a delegated task, the skill that matters is the one good managers already have. Ethan Mollick makes this case well in his piece on management as an AI skill: as agents take on work that would take a person hours, the value of being able to delegate clearly goes up. Instructing an agent on a real task starts to look like the documents firms already use to hand work off, a scope, a brief, a set of deliverables. What are we trying to do and why, where the limits are, what "done" looks like, and what to check before calling it finished.

Mollick's point is that these are management basics, not technical tricks. If your people can explain what they need, give useful feedback and judge whether the output is any good, they can work with an agent. The scarce skill is knowing what to ask for and recognising a good answer, which is precisely what an experienced professional has and a clever prompt does not.

This is the work we do with leaders: taking the tasks a team already runs and turning them into clear, repeatable briefs an agent can act on, so the gain is real and the judgement stays with a person. The tools change every few weeks. The discipline of delegating well does not.

Where a human still has to stay in charge

Faster delegation raises the stakes on checking. An agent that works for an hour can produce an hour's worth of plausible, confident, wrong. In regulated or client-facing work, a mistake that reads well is more dangerous than one that looks obviously off.

So the rule that makes this safe is dull and firm: every delegated task has a named human who owns the output and checks it before it goes anywhere. Not a glance, a real read against what good looks like. Some tasks should not be handed over at all, the ones where the judgement is the job rather than the grunt work around it. We have written separately about what to keep off the delegation list, and the line holds here. The agent drafts and assembles. The person decides and signs.

What to do this month

Pick one non-technical function and one real task. Recruitment longlists, a first-pass finance summary, a research pack, whatever takes an hour by hand and happens often. Write a proper brief for it, run it through an agent a few times, and have the person who owns that work judge the output honestly against what they would have produced. You will learn quickly whether it saves real time and where it needs a human eye. That single loop teaches more than another tool trial.

If it would help to do that with someone who has run this with other firms, our AI Lessons for Leaders coaching is built around exactly this: your real tasks, turned into workflows your people can trust, with judgement kept in charge. A good starting point is a one-to-one discovery call.

Frequently asked questions

Are AI agents only useful for technical teams?

No. The clearest recent evidence, from OpenAI's year of Codex data, is that non-developer use grew faster than developer use, and that legal, finance and recruiting teams adopted agents heavily once the tools could handle their work. Agents started as a coding aid and have spread well beyond it.

What kinds of tasks can non-technical staff hand to an AI agent?

Work that takes an hour or more by hand, has a clear goal, and produces something you can check: a first-pass market map, a draft variance summary with workings, a cited research pack, a structured longlist from messy notes. The agent assembles a draft; a person refines and decides.

Do you still need a human if the agent does the work?

Yes, and more so as the tasks get longer. Every delegated task should have a named owner who checks the output against what good looks like before it is used, especially in regulated or client-facing work. Some judgement-heavy tasks should not be delegated at all.

Is the OpenAI research reliable?

Treat it as useful but interested. OpenAI is reporting on its own product and, for the internal figures, its own staff, and the task-time estimates are model-generated on a small sample, so they are directional. The trend it shows, non-technical adoption rising fastest and tasks getting longer, is the part to act on.

Sources

• OpenAI, How agents are transforming work, 25 June 2026: https://openai.com/index/how-agents-are-transforming-work/
• Ethan Mollick, Management as AI superpower, 27 January 2026: https://www.oneusefulthing.org/p/management-as-ai-superpower