AI Is Everywhere at Work. Almost Nobody Is Letting It Finish the Job.

Fourteen million conversations. One hundred fifty countries. Here's what Google learned when it actually counted how people use AI at work: everyone is using it, and almost nobody is letting it finish the job.

The featured story

Google's first ATLAS study tracked 14.65 million real Gemini interactions over two weeks in April. The picture is broad but shallow.

AI showed up in 68% of occupations — work that covers 88% of U.S. employment. But inside the median occupation, AI touched only about 21% of the tasks. And in the non-routine cognitive work where AI is used most, fewer than 10% of conversations tried to automate a task end to end.

The rest is assistance: research, drafts, troubleshooting, summaries, strategy. People are using AI as a brilliant intern, not a replacement worker.

That's not a failure. It's a gap. The 90% who use AI to assist, and the 10% who wire it to finish, are heading to different places.

Quick hits

CIOs are all in — and a little scared. Logicalis's 2026 CIO Report: 60% of global CIOs plan agentic AI investment in the next 12 months, and 53% expect generative and agentic AI to disrupt their business model within two years. IDC projects more than one billion AI agents operating globally by 2029, processing 217 billion actions a day.

Agent governance just became a product. Nuggets launched the Authority Control Plane — software that sits between an agent and your systems and blocks unauthorized actions in real time, with an estate-wide kill switch. Their CCO: "Customers told us they can't move agents to production because they can't prove what an agent was allowed to do."

The bluntest quote of the week. SOFTSWISS CTO Sergey Kastukevich: basic AI adoption delivers just a 10-15% lift unless you overhaul the process underneath. "If you put a great engine in the old car, it will not work." That's Wednesday's deep dive.

Tool of the week

Microsoft Intelligent Terminal 0.2. AI agents that live inside your terminal — per-tab agent selection, OpenCode integration, WSL support, and local models. Ten minutes to set up, and it quietly removes the copy-paste tax from your workflow. Free.

Tomorrow (Wednesday): why your AI only returns 10% — and the five-step workflow redesign that fixes it.