Which departments are adopting AI coding tools fastest?

Per OpenAI's Enterprise Signals data, legal grew 108× in weekly active enterprise Codex users since February, followed by sales and recruiting at 41× each and marketing at 26×. Engineering grew 5× — the slowest of the group.

What separates high-adoption companies from typical ones?

Connection, not conversation. Frontier firms — the top 10% by usage — show 21% of weekly active users on Plugins against 9% at typical firms, and 19% on skills against 3%. They also generate 8.3× the output tokens per active user, up from 2.6× in January.

How reliable are these numbers?

They are OpenAI's own figures about OpenAI's own enterprise customers, published by OpenAI. That does not make them wrong, but it does make them selected. Cite them as vendor data rather than independent research.

Codex Is Now 64% of Enterprise Output

A traditional law library, one glowing tablet open on the reading table among bound volumes — DataAugust 12, 2026·2 min read·By the Automation Squad Research
Codex Is Now 64% of Enterprise Output
If you're selling automation into non-technical departments, this is the proof slide.

By Robert MacKelfresh

Founder, Automation Squad · August 12, 2026

The short answer

OpenAI's Enterprise Signals report, published August 12, 2026, finds Codex generated 64% of combined Codex and ChatGPT output tokens among enterprise customers as of June. Since February, weekly active enterprise Codex users grew 108× in legal, 41× in sales, 41× in recruiting and 26× in marketing, against 5× in engineering.

Adoption table

The numbers, and the one that changes who you sell to

All figures are OpenAI's own, about OpenAI's own customers. Read them with that in mind — and then read the growth column anyway, because the ordering is the story.

Notice which departments are missing from your own rollout
If your AI programme is aimed at engineering, the growth ordering says you are aiming at the slowest-growing group. Legal grew twenty times faster than engineering off a much smaller base — small bases grow fast, but 108× against 5× is not explained by base effects alone.
Check whether your organisation is pulling away or falling behind
The frontier-versus-typical gap widened from 2.6× to 8.3× in five months. Whatever compounding is happening, it is accelerating, which means the cost of waiting is not flat.
Look at the Plugins and skills numbers as a diagnostic
21% versus 9% on Plugins, 19% versus 3% on skills. The separator is not how much people chat — it is whether they have connected the model to anything. That is a fixable gap and a cheap one.
Pick your least technical high-volume department and run one pilot
Legal, recruiting and sales are where the growth is. Choose the single most repetitive document-shaped task in that department and give it a proper four-week trial rather than a demo.
Read the source's provenance before you quote it in a deck
This is OpenAI reporting on OpenAI's customers. The numbers are specific and probably accurate; they are also selected. Cite them as vendor data, because someone in the room will check.

Measure	Figure
Codex share of enterprise output tokens (June)	64%
Codex WAU growth since February — legal	108×
— sales	41×
— recruiting	41×
— marketing	26×
— engineering	5×
Frontier firms vs typical, output tokens per active user	8.3× (was 2.6× in January)
Weekly active users using Plugins — frontier vs typical	21% vs 9%
Weekly active users using skills — frontier vs typical	19% vs 3%
OpenAI's own employees using Plugins weekly	95%

Take it with you

OPENAI ENTERPRISE SIGNALS — Aug 12, 2026
Source: https://automationsquad.com/news/openai-enterprise-signals-adoption/

THE HEADLINE
Codex = 64% of combined Codex+ChatGPT enterprise output tokens (June)

GROWTH IN WEEKLY ACTIVE ENTERPRISE CODEX USERS, SINCE FEBRUARY
  Legal ............ 108x
  Sales ............ 41x
  Recruiting ....... 41x
  Marketing ........ 26x
  Engineering ...... 5x     <- the slowest

THE GAP IS WIDENING
  Frontier firms vs typical, output tokens per active user
    January ... 2.6x
    Now ....... 8.3x

WHAT SEPARATES THEM (it isn't chat volume)
  Using Plugins ... frontier 21%  vs typical 9%
  Using skills .... frontier 19%  vs typical 3%
  OpenAI's own staff using Plugins weekly ... 95%

WHAT TO DO
[ ] Which departments are missing from my rollout?
[ ] Am I aiming at engineering — the slowest-growing group?
[ ] Have my people CONNECTED anything, or just chatted?
[ ] Pick one non-technical, high-volume, document-shaped task
[ ] Run a 4-week trial, not a demo

PROVENANCE
Vendor data: OpenAI, about OpenAI's customers. Specific, probably
accurate, definitely selected. Cite it as such.

OpenAI published Enterprise Signals on August 12, 2026, alongside a working paper titled How Organizations Use AI: Evidence from ChatGPT. It is one of the few places any lab publishes usage data at this granularity.

The facts: as of June, Codex generated 64% of combined Codex and ChatGPT output tokens among enterprise customers. Frontier firms — the top 10% by usage — generate 8.3× as many output tokens per active user as typical firms, up sharply from 2.6× in January. At those frontier firms 21% of weekly active users use Plugins against 9% at typical firms, and 19% use skills against 3%; inside OpenAI itself, 95% of employees use Plugins weekly. The most striking figure is the growth ordering: since February, weekly active enterprise Codex users grew 108× in legal, 41× in sales, 41× in recruiting and 26× in marketing, against 5× in engineering.

Automation Squad's take: a coding tool growing twenty times faster in legal than in engineering tells you the category name is wrong. What legal, recruiting and sales have in common is not code — it is repetitive, structured, document-shaped work with a clear right answer, which is exactly what these tools are good at and exactly what nobody built a product page for. The second finding is quieter and more useful: the difference between the leading firms and everyone else is not that their people chat more, it is that their people have connected the model to something. Plugins and skills adoption separates them by two to six times. That is a fixable gap, and fixing it costs an afternoon rather than a budget cycle.

Run this now: list the departments in your organisation and mark which ones your AI rollout has actually reached. If the answer is engineering and nobody else, you are aimed at the slowest-growing group in OpenAI's data. Pick your most document-heavy non-technical function, find its single most repetitive task, and run a proper four-week trial rather than a demo. And check the connection question specifically — ask five people whether they have ever connected the model to a real system, because that answer predicts more than any amount of training.

By the numbers

The data behind the story — figures from cited sources, with our own analysis labelled.

Codex weekly active enterprise user growth since February, by functionPer OpenAI — Enterprise Signals, August 2026

Legal	108×
Sales	41×
Recruiting	41×
Marketing	26×
Engineering	5×

The ordering is the finding. Engineering is last, by a wide margin, in adoption of a coding tool.

Questions people are asking

Which departments are adopting AI coding tools fastest?: Per OpenAI's Enterprise Signals data, legal grew 108× in weekly active enterprise Codex users since February, followed by sales and recruiting at 41× each and marketing at 26×. Engineering grew 5× — the slowest of the group.
What separates high-adoption companies from typical ones?: Connection, not conversation. Frontier firms — the top 10% by usage — show 21% of weekly active users on Plugins against 9% at typical firms, and 19% on skills against 3%. They also generate 8.3× the output tokens per active user, up from 2.6× in January.
How reliable are these numbers?: They are OpenAI's own figures about OpenAI's own enterprise customers, published by OpenAI. That does not make them wrong, but it does make them selected. Cite them as vendor data rather than independent research.
What share of enterprise AI usage is Codex?: As of June, Codex generated 64% of combined Codex and ChatGPT output tokens among enterprise customers.