Your agent is giving you reasonable answers from half your context. A conversation with OpenAI.

When I asked which teams inside OpenAI had crossed over after software engineering, legal came up early.

In my conversation with Andrew, who leads the team behind OpenAI’s Codex desktop app, and Akshay, who leads OpenAI’s productivity and engineering work, they described different functions reaching a point where AI became useful for much more of their jobs. Coding came first, because everything was already a local CLI. Legal was next, because the material was sitting in documents the model could open. Other functions crossed over as access to their information improved and the tools became better at using it.

Technical ability was part of the explanation. So was something more ordinary: could the AI actually get to the material a person needed to do their work?

We talked about people in finance building applications, colleagues creating tools for other colleagues, and what happens when making things gets easy enough that deciding what to make becomes a much larger part of the job. And a question I have been asking myself: when I can give so much of the busywork to an agent, what am I doing with my time?

Those questions belong together.

My view is that the most consequential change comes when something one person learns improves what other people can do. AI creates more ways for that to happen. Whether it does happen depends on the environment people work in, the authority they have, and what the organization recognizes as valuable work.

I spend a lot of time building with AI and talking with people trying to make it useful inside organizations. A recurring problem is that a company can have some extraordinarily capable AI users and still struggle to become more capable as a company. One employee is building tools and delegating analyses, and their colleagues have seen enough to want similar results. Yet the work around that person still moves through the same bottlenecks, and everyone keeps asking how to get more people to work like them.

OpenAI is an unusual place to look for lessons about AI adoption. Its people build the products, work close to the research, and have unusually strong reasons to use what they are making. Andrew and Akshay’s account is one sample. What makes the conversation valuable is the opportunity to inspect habits that are forming close to the technology and consider which ones might travel.

This essay covers:

• Before you judge adoption, look at the environment. A person’s experience of AI depends partly on whether the system can reach the information their job requires.

• How individual expertise becomes useful to a team. Most of it is still available only through repeated access to the person who has it.

• What onboarding has to teach. An accurate sense of what the tool cannot reach, and how it behaves when the material is missing.

• How leadership changes when more people can build. More possibilities become visible, and somebody has to decide which ones deserve attention.

• How we develop judgment when AI can do work we once learned by doing ourselves.

• Who actually gets the time back. The capacity someone creates might support a more interesting responsibility, or it might become a higher production target.

Start with the environment. Before you decide who on your team is resistant, find out what their AI can actually reach.