People quietly use AI their own way at work — but is that the same thing as the company's official AI rollout?

This explores the gap between personal, ground-level AI use at work and the official version a company sanctions and measures. The corpus suggests these aren't two versions of the same thing. They run on different logics, and the distance between them is where much of the real story sits.

Start with who holds control. When companies deploy AI officially, the organization decides first and the individual decides second. A study of junior and senior engineers found that tool mandates, allow-lists and data policies set how much control engineers keep over agentic AI before personal preference comes into play at all Does personal preference shape how engineers use AI tools?. Inside those limits, newer engineers swung between leaning on the tools too much and avoiding them. That is a coping problem the policy itself doesn't address. Individual use, by contrast, starts from personal curiosity and convenience. The cost is that it has no guardrails and no shared practice.

Next, who actually uses it. OpenAI's own telemetry from 1,764 firms shows official adoption clustering in larger, R&D-heavy companies. Even inside those companies, use is uneven: marketing staff and early-career workers send far more messages than executives and senior people Who adopts enterprise AI first and how do they use it?. So the people who approve a deployment are often not the people who use it most. A separate line of research tracks where work has been formally handed to AI inside structured workflows. That kind of delegation concentrates in information-heavy jobs and follows what the technology can actually do, not how popular chatbots are Where have workers actually delegated tasks to AI?. Casual chatting with an AI and building it into a workflow are different kinds of adoption, and they spread along different paths.

This is the part you might not expect: individuals often hide their use. Across four experiments with more than 4,000 people, AI users expected colleagues to see them as less competent and less diligent. They were also less willing to tell managers they had used it Do people fear judgment when they use AI at work?. That points to a blind spot in official numbers. Telemetry from a sanctioned tool can't see personal use that people keep quiet, so the company's picture of its own AI use may miss part of what is really happening, especially among the junior staff who use it most.

Finally, closing the gap from either side is harder than it looks. Benedict Evans argues that making tools easier to build doesn't fix two deeper problems Does easier tool-building actually solve enterprise adoption problems?. First, most workers don't recognize their own tasks as things AI could automate. Second, company-wide adoption needs decisions that cut across departments and budget cycles. An individual can work around both by tinkering alone. An organization has to work through them. That is why personal use can move fast and stay invisible, while official deployment moves slowly and is shaped by policy before anyone opens the tool.