Since February, AI use inside corporate legal teams has grown 108 times over. In engineering, over the same period, it grew five times.
That comes from OpenAI’s two enterprise adoption reports, published on 12 August. It is worth sitting with, because it inverts the assumption most organisations are still working from: that AI adoption is a technology story, that it happens fastest among technical people, and that the rest of the business will catch up later.
The rest of the business is not catching up later. In several functions it is already in front.
The number underneath the headline
OpenAI splits its enterprise customers into two groups. Frontier firms are the top 10 per cent, measured by how much work their people push through the tools. Everyone else is typical.
In January, frontier firms were getting 2.6 times as much through per person as typical firms. By June that had reached 8.3 times.
The gap roughly tripled in seven months. That is the figure worth writing down, because it describes a rate rather than a state. A stable gap is a competitive fact you can plan around. A gap widening at this speed is a different kind of problem, and the thing about compounding differences is that they look survivable right up until they do not.
What the measure actually counts
The unit is output tokens per active person, and it needs translating.
A token is a fragment of text. Output tokens are roughly how much material the model produced for somebody. So this measures the volume of work being pushed through the tools, not the value coming back out. A company can run enormous volume and get very little for it. I will come back to that on Friday, with a company that did exactly this and said so publicly.
Read 8.3 times as intensity of use. It is a real signal, and it is not the same as return.
One further figure sets the context. As of June, Codex generated 64 per cent of the combined Codex and ChatGPT output tokens across enterprise customers. Codex is the product that carries out tasks rather than answering questions. Most of the volume has moved to tools that do work, which is the shift the reports are named for: assistance to execution.
Where the growth actually is
Growth since February, by function:
- Legal: 108 times
- Sales: 41 times
- Recruiting: 41 times
- Marketing: 26 times
- Engineering: 5 times
Engineering is last by a wide margin, and there is a plain reason. Engineering started early. It has been using these tools seriously for two years, so its base was already high and its growth rate is correspondingly low. Legal grew 108 times partly because it was starting from close to nothing.
That is the honest reading and it does not weaken the finding. It sharpens it. The functions that spent two years watching from the sidelines have moved, and they have moved quickly.
Frontier firms are also using more of what the tools offer. 21 per cent use plugins and 19 per cent use skills, against 9 per cent and 3 per cent at typical firms. On skills that is more than six times the adoption rate. It is not a licence gap. It is a practice gap, and practice gaps close slowly because nobody can buy their way out of one.
The finding nobody wants on a slide
Six months after adoption, early-career employees send 13 more messages a week than executives.
The easy reading is that senior people are slower. I do not think that is quite it. Executives have assistants, delegated work and calendars they do not control. Somebody three years into their career has none of that, and every incentive to find an edge.
The problem is what it implies about where the knowledge now sits. The people with the least authority to change how work gets done are the ones who best understand the tool that could change it. That is a management problem rather than a training problem, and another round of licences will not touch it.
What this is not
This is OpenAI reporting on adoption of OpenAI products. Every figure comes from a company with a direct commercial interest in the conclusion, and none of it has been independently verified.
Treat the multiples as directional. Growth of 108 times on a small February base is not the same as large absolute volume, and anyone presenting that number without saying so is selling something. The shape of the finding is more reliable than its precision: non-technical functions are moving faster than technical ones, and the leaders are pulling away from the middle.
The customer example carries the same caution. OpenAI reports Virgin Atlantic completing engineering refactors in 30 minutes rather than two weeks, and product teams finishing weeks of competitive research in hours. Those are Virgin Atlantic’s own accounts, published by their supplier.
The question to take into the week
If you take one thing from this, do not make it the 8.3 times.
Make it the ranking. Legal, sales, recruiting and marketing are all ahead of engineering, and that is the reverse of what most organisations have planned for. The budget, the pilot and the internal expertise usually sit with the technical team, because that is where everybody assumed this would happen first.
So ask which function in your organisation has actually moved furthest. If the answer is engineering, look again. Either you are measuring the wrong department, or you have a legal team quietly doing something nobody has written down.
Tomorrow I want to look at the other half of this: what separates the people getting real value from these tools from the people getting very little. It is not the licence, and it is not the budget. It turns out to be the expertise the person already had.
Source: From assistance to execution: How enterprises put AI to work, OpenAI, 12 August 2026. All figures are published by OpenAI and describe adoption of OpenAI products.