The Top 10% of Companies Now Produce 8.3x More AI Output Per Person. In January It Was 2.6x.

In January, the top 10% of enterprise AI users produced 2.6 times as much AI output per person as a typical company. By August that figure was 8.3 times.

The gap tripled in seven months. That is the single most important number in enterprise AI right now, and it is not being talked about nearly enough, because the story it tells is uncomfortable: this is not a race where everyone is slowly catching up. It is a race where the leaders are accelerating away.

The data comes from OpenAI's Enterprise Signals report, published August 13, which measured output tokens per active user across its enterprise customer base and defined frontier firms as the top decile of AI usage each month.

What that number actually measures

Output tokens per active user is a better metric than most of what gets reported, and it is worth being precise about why.

It does not count seats purchased. It does not count logins, or pilots launched, or the percentage of staff who have "access to AI." Those are the numbers that fill vendor decks, and they measure procurement, not work.

Output tokens per active user measures how much was actually produced, per person who was actually using it. A company where everyone has a licence and nobody does anything with it scores badly, correctly. A company where a smaller number of people are pushing enormous volumes of real work through agents scores well, also correctly.

So an 8.3x gap is not 8.3 times more enthusiasm. It is 8.3 times more work coming out per active person.

It is not model access

Here is what makes this interesting rather than merely dramatic.

Every company in that dataset has access to the same models. There is no frontier tier of GPT that typical firms cannot buy. The gap did not open because leaders bought better intelligence, because they did not.

The report points at what does differ, and it is unglamorous. Frontier firms adopt the capabilities that connect AI to their own systems. 21% of active users at frontier firms use plugins, against 9% at typical firms. They wire agents into company context, actual tools, and repeatable workflows, rather than leaving them as a very smart text box that knows nothing about the business.

That is the entire difference. Not a better model. A model that has been given the company's context, permission to use real tools, and a defined job that recurs.

We have been making this argument all year from first principles. It is satisfying, and slightly alarming, to see it show up as a tripling gap in seven months of real usage data.

The part almost nobody has noticed

Buried in the same report is a finding that reframes who this is about.

Since February, weekly active enterprise users of agentic coding tooling grew 108x in legal, 41x in sales, 41x in recruiting, and 26x in marketing. In engineering, it grew 5x.

Read that again. The tools built for developers are growing twenty times faster outside engineering than inside it.

Legal teams are not writing software. They are using agentic tooling to work through contracts, extract obligations, compare versions, run repeatable multi-step processes over documents. Sales teams are doing research and outreach sequences. Recruiters are processing pipelines. The capability that got built to help engineers execute multi-step work turns out to be a general-purpose execution capability, and the functions with the most repetitive structured work are adopting it hardest.

The relevant number for the token mix backs this up: as of June, agentic tooling generated 64% of combined enterprise output tokens, against conversational use. Enterprise AI has already tipped from chat to execution. Most companies' internal policy conversations have not.

Why this gap compounds instead of closing

Ordinarily you would expect a gap like this to narrow. The laggards catch up, best practice spreads, tooling commoditises.

This one is widening, and the mechanism is worth understanding if you are on the wrong side of it.

Workflow design is a learned skill, and the learning is cumulative. A company that has connected fifteen processes to agents is dramatically better at connecting the sixteenth than a company doing its first. They know which processes decompose well, where a human check belongs, what failure looks like.

Context accumulates too. Every month an agent operates inside your business, it acquires more of your specifics: your clients, your terminology, your exceptions, your standards. A company two years into that has an asset a competitor cannot buy, because it was not for sale. It was accrued.

And the organisational habit compounds hardest of all. In frontier firms, reaching for an agent is the default reflex when a repetitive task appears. In typical firms it is still a project that needs a business case. The reflex is worth more than the technology.

Three compounding curves at once. That is why 2.6x became 8.3x rather than 2.6x becoming 2.1x.

What to do if you are on the wrong side

The gap is by industry as well as by company. Information and technology shows the widest spread at 11.7x, manufacturing the narrowest at 5.3x. Even in the sector with the smallest gap, the leaders are producing five times the output per person. There is no industry where this is not happening.

Three things, in order.

Stop measuring adoption by seats. Count output per active user, or at minimum count how many defined processes an agent currently runs end to end. If that number is zero, your adoption rate is zero, whatever the licence report says.

Connect it to your actual context. The single largest divergence in the data is between AI that knows nothing about your company and AI wired into your systems, tools and history. An assistant that cannot see your data is a demo. An agent that remembers your clients, follows your process and holds your context is a colleague.

Pick something repetitive and give it away completely. Not a pilot, not a trial. One recurring process, handed over end to end, with a named person accountable for the result. That is the unit that compounds. Capabilities do not compound. Delegated processes do.

The companies pulling ahead did not get access to better intelligence than you have. They started earlier on the boring part: giving their agents context, tools and real jobs. Seven months ago that was worth a 2.6x edge. Today it is 8.3x, and the curve is not flattening.

Want to test the most advanced AI employees, with the memory and context to actually do your work? Try it here: https://geta.team