The most conservative white-collar industries are turning out to be the most aggressive early adopters of coding agents. Which is not how anyone called it.
Inside OpenAI, the lawyers, finance team, and recruiters now use Codex, the company's coding agent, as their primary AI tool. According to OpenAI's own breakdown, legal, finance, and recruiting staff crossed over around April and now generate more than 85% of their output tokens through the tool. The median lawyer at the company is running 12 times more work through Codex than they did in November.
That's a strange sentence. Lawyers. Tokens. Codex.
The pattern shows up outside OpenAI too. Hedge fund Walleye Capital says 100% of its roughly 400 employees use Claude Code, Anthropic's equivalent of Codex. Citadel has its analysts using Claude inside Excel to build coverage models and separate signal from noise. And Big Law firms are running internal "vibe coding" programs where non-engineers spin up custom tools for case prep.
Finance and legal are usually the last industries to touch anything that affects workflow. Both are heavily regulated, both bill by the hour, and neither has historically rewarded engineers in suits. So watching them pull coding agents deeper into the daily routine than some engineering teams is the genuinely surprising part.
What's actually happening is that "coding agent" turned out to be a misnomer. People aren't using these tools to ship production software. They're building one-off scripts, automating spreadsheet work, restructuring documents, and standing up internal tools that used to require begging an engineering team for time. About a quarter of the Codex work done by non-engineers is coding-shaped. The rest is research, drafting, and analysis.
Romain Huet, OpenAI's head of developer experience, said on a recent podcast that walking around the office, "every few minutes you would hear something like 'just ask Codex.'" His own read is that people are still underusing it. "The models are so good that we're still too shy with them."
Not everyone is sold on the trend. Chris Bedi, ServiceNow's Chief Customer Officer, has been openly skeptical of measuring AI progress by how many tokens a company is burning. "It's almost like measuring a restaurant based on how many ingredients they buy," he said. "You don't measure a restaurant that way." His more practical warning is that somebody eventually pays for all those tokens.
There's a quality concern too. Stanford's legal-tech researchers found that teams using agents without domain-specific verification introduced 12-18% more compliance revisions than traditional methods. The lawyers vibing their way through case prep are also generating extra cleanup somebody else has to catch.
INTO THE VALLEY
The expected story about AI in the enterprise was that engineers would adopt first and everyone else would follow years later. What's happening instead is that the suits are running the same tools as the engineers, just for entirely different reasons. Finance and legal didn't fall for Codex because they want to ship software. They fell for it because the agent lets one person do work that used to require asking three other departments. If that pattern holds, the industries most exposed next are the ones built around middle layers nobody wants to call anymore.
