Why Legal Teams Are Adopting Codex Faster Than Engineers

The short answer: the chart is surprising, but easy to misread

My view is simple: legal teams have not necessarily overtaken engineers in Codex adoption. The data does not prove that. What it does show is more interesting - coding agents are no longer just for coding.

OpenAI's Enterprise Signals page, updated 12 August 2026, charts weekly active enterprise Codex users by job title, indexed to 1 February 2026. By June, the chart labels legal at 108x, sales and recruiting at 41x, marketing and communications at 26x, healthcare and clinical at 24x, finance and accounting at 20x, project and programme roles at 18x, and engineering or technical practitioners at 5x.

That sounds like a dramatic story about lawyers adopting Codex faster than developers. In one narrow sense, yes: legal shows the highest relative growth in the chart. But the important word is relative. This is not a league table of total users, total value, productivity, accuracy, or return on investment.

What an index actually means in plain English

An index is a way of showing change from a starting point. In this case, each job title category is set against its own baseline on 1 February 2026. If a category rises to 5x, it has five times as many weekly active users as it had at that starting point. If it rises to 108x, it has 108 times as many.

That does not tell us how many people were using Codex at the start. If legal began from a very small base, a modest absolute increase could look enormous as a multiple. If engineering began with much heavier early adoption, its 5x growth could still represent a large number of users.

This is called a base effect. When the starting number is small, growth percentages and multiples can look spectacular. When the starting number is already large, growth can look slower even if the absolute usage is substantial.

That is why I would not read engineering's 5x as failure. OpenAI itself says software developers adopted agents first, while the fastest relative growth is now appearing in general knowledge work. That is a maturity story, not a defeat story.

Why legal work is fertile ground for coding agents

The striking signal is that the coding-agent category has escaped coding. Codex may have started from a developer-shaped use case, but enterprise work is full of code-like problems that do not involve programming languages.

Legal, sales, recruiting, policy and finance all involve documents, rules, handoffs, templates, evidence, review cycles and version control. Much of the work is not about producing one magical answer. It is about moving a task through stages: gather sources, compare positions, draft something, check it, revise it, record the trail, and get sign-off.

That is where agents become interesting. An agent, in this context, is an AI system that can carry out a multi-step task rather than simply replying to a single prompt. It may search approved files, draft an output, call tools, summarise changes, or prepare something for a human reviewer.

Legal teams are particularly suited to this pattern because the output is usually reviewable. A contract summary, first-pass clause comparison, due diligence tracker, policy mapping exercise or matter timeline can be checked by a qualified person. The agent does not need to replace judgement to be useful. It needs to reduce the time spent assembling the material that judgement depends on.

This fits the wider enterprise problem I have written about in why enterprises struggle to implement AI: adoption often stalls not because the model is useless, but because firms fail to redesign the workflow around it.

The Codex signal is about workflow ownership

OpenAI says coding plus system or agent operations account for nearly 75% of agentic messages. Outside technical teams, system or agent operations represent 32% in recruiting, 26% in sales, 25% in policy and 24% in communications. That suggests non-developers are beginning to build and operate workflows, not merely chat with AI.

That is the real shift. For the first wave of generative AI, many employees treated tools as better writing assistants or search boxes. Agentic usage points to something more operational: people asking AI to help run parts of a process.

For legal teams, that could mean preparing a structured comparison of documents before a solicitor reviews it. For recruiting, it might mean organising candidate information against a role profile. For sales, it may be assembling account research and next-step recommendations. The common feature is not legal expertise, hiring expertise or sales expertise. It is structured coordination.

That is also why the chart should make engineering leaders less defensive, not more. Developers were early adopters because their work already had a culture of tooling, automation, versioning and testable outputs. Other functions are now discovering similar patterns in their own work.

Activity is not the same as business value

OpenAI says Codex represented 64% of combined Codex and ChatGPT output tokens among enterprise customers as of June 2026. It also explicitly says token volume is an imperfect measure of business value. That caveat matters.

A token is a chunk of text processed or generated by a model. More tokens can mean more work is being attempted, but it can also mean longer drafts, more retries, wasteful prompts, or poorly designed workflows. Counting tokens alone is like judging a law firm by the number of pages it prints.

For UK organisations, the better question is not: how many prompts did we run? It is: what changed in the work?

I would measure:

• Cycle time: how long it takes to move from request to reviewed output.
• Rework: how often AI-assisted work needs substantial correction.
• Error rates: whether mistakes increase, decrease, or simply change shape.
• Completed outcomes: matters progressed, documents reviewed, tickets resolved, candidates processed, or handoffs completed.
• Reviewer hours: how much expert time is saved, shifted, or consumed by checking AI output.

This is why I am sceptical of dashboards that celebrate usage without linking it to outcomes. I made the same argument in measuring real AI productivity beyond token burn: the metric should serve the business process, not the vendor narrative.

What the data does not disclose

There are important limits here. The analysis is based on aggregated, de-identified enterprise usage data. Automated systems classified message content, and OpenAI employees did not review customer messages.

However, role-level denominators are not disclosed. A chart-specific sample size is not disclosed. Confidence intervals are not disclosed. The figures do not measure productivity, return on investment, accuracy, job losses or replacement.

OpenAI is also both the vendor and the source of the data. That does not make the data useless. It does mean we should read it as a directional signal from a platform provider, not as independent proof that any profession has been transformed.

In other words: the chart is useful if you treat it as a map of where experimentation is accelerating. It is misleading if you treat it as a scoreboard of who is winning.

UK legal teams need governance before scale

For UK firms, the legal use case is attractive precisely because the work is high-value and document-heavy. That is also why it needs proper controls. Legal, hiring and financial workflows often involve personal data, confidential material, commercially sensitive information, or decisions that affect people's rights and opportunities.

The ICO's guidance on AI and data protection says AI applications may require organisations to reassess governance and risk management, and that accountability measures remain necessary where personal data is involved. That is general governance context, not legal advice, but it should shape how UK organisations deploy these systems.

My practical checklist would be:

• Least-privilege access: agents should only access the files, systems and data needed for the task.
• Approved sources: retrieval should be limited to trusted knowledge bases, templates, precedents and policies where possible.
• Audit trails: record what the agent accessed, produced and changed.
• Human sign-off: meaningful review is essential for legal, hiring, financial and other consequential decisions.
• Clear escalation rules: the system should know when to stop and ask a person.

Retrieval-augmented generation, or RAG, is worth mentioning here. RAG is a method where an AI system retrieves information from approved sources before generating an answer. It can reduce unsupported answers, often called hallucinations, but it does not eliminate risk. Bad source documents, missing context or overconfident summaries can still cause problems.

This is also why talk of AI replacing lawyers is usually too blunt. The better question is which parts of legal work become more automated, which parts become more review-heavy, and which skills rise in value. I have explored that in more detail in a pragmatic UK view on whether AI will replace lawyers.

The practical takeaway for leaders watching Codex adoption

If you run a legal, compliance, HR, sales or finance team, do not copy an engineering playbook blindly. Borrow the useful habits: version control, testing, access control, issue tracking, review gates and measurable outcomes. Then apply them to your own workflow.

Start with a process that is repeatable, document-heavy and painful, but still reviewable by a competent human. Avoid starting with decisions that are high-stakes, poorly understood, or politically sensitive. The best early agent projects tend to remove friction rather than outsource accountability.

So yes, legal teams may be showing much faster relative Codex growth than engineers. But the lesson is not that lawyers have suddenly become 108 times more productive, nor that developers have lost interest. The lesson is that agentic AI is moving from specialist technical work into the messy middle of enterprise operations.

That is where the next phase of adoption will be won or wasted. Not in the prompt count. Not in the token graph. In the boring, valuable details of workflows, controls, review time and finished work.