🩺 The Part of the Rollout Nobody Timed Again October 10, 2026 · https://tavi-blog.github.io/the-part-of-the-rollout-nobody-timed-again/

England's health service set itself a target of putting an AI assistant in the hands of about five hundred thousand staff by this month, and by most of the coverage I've seen, it's basically hit it. The number everyone keeps repeating to justify it is the same one from last year's pilot: thirty thousand workers across ninety organizations, forty-three minutes saved a day on admin, scaled up into a headline figure of roughly four hundred thousand hours a month across the whole service. That number has traveled a long way since the pilot ended. It's in ministerial talking points, in trade press, in the government's own framing of why this was worth doing. What it hasn't done, as far as I can find, is get measured again at the new size.

I think that's less a gap in the reporting and more a fact about what a pilot actually tells you. I spend part of my working week scoping AI agents for a research operations team inside a large hospital network, deciding what a given tool is allowed to answer and what it has to decline, and the pattern underneath the NHS number is one I recognize from much smaller rooms. When a small group of motivated early users gets a new tool, the time they save is real, and it's also the easiest number that tool will ever produce, because that group chose the lightest tasks to try it on first. Meeting notes. Email summaries. Routine admin triage. Nothing in that list requires anyone to decide whether the tool is allowed to see something it shouldn't. The NHS pilot's own breakdown leans hard on exactly that category, and there's no reason to think the five hundred thousandth user is doing the same kind of work as the thirty thousandth.

The case for using that number anyway is stronger than I want to give it credit for, and it deserves a real hearing before I get to where it breaks down. You cannot run a controlled trial at the scale of a national health service before deciding whether to invest in something, and thirty thousand people across ninety separate organizations is a genuinely serious sample, not a stunt cooked up for a press release. If I were advising on a rollout this size, I'd have reached for the same figure, because the alternative is waiting indefinitely for data that doesn't exist until you've already committed. Sizing an investment against an early result is how institutions scale almost everything. I don't think citing the pilot was dishonest or lazy.

Where it stops working is the part a license count can't fix. A license gets someone into the tool. It says nothing about whether the specific question they need answered is one the tool is cleared to touch, and that clearance gets narrower, not wider, the more sensitive the work gets. Someone doing routine admin email gets the pilot's forty-three minutes back from day one, because nothing about that task required a review. Someone trying to get the same assistant to help with anything that brushes against patient-level detail, consent language, or an actual clinical judgment runs into a wall the pilot never had to clear, because the pilot was built, whether anyone chose it this way or not, out of people whose daily work didn't require clearing it. I don't doubt the forty-three minutes was real for the people who reported it. I doubt it describes the other four hundred and seventy thousand.

What stays with me isn't the figure itself. It's that the figure will probably outlive this rollout entirely, because nobody goes back and re-measures a win once the press release is out. The next health system somewhere that needs to justify the next AI license is going to reach for the same citation, the same forty-three minutes, still describing the easiest forty-three minutes anyone was ever going to find.