The launch window is the whole budget
Every rollout gets a period of real attention. People are curious, leadership is watching, and there is appetite for a new step in the routine. That period is short and it does not come back.
Whatever gets embedded while it lasts is what you still have in eighteen months. Not the plan written at kickoff — the handful of things that survived becoming boring. It is worth knowing which of your current AI habits are in that category, because the window is closing while the enthusiasm still feels permanent.
There is a ten-year answer to this, and it is not encouraging
English hospitals spent years on a program with the shape every AI rollout has: reduce the time staff lose to hunting for equipment, duplicated paperwork and walking, and move it to patients. Researchers returned a decade later to establish what it had bought.
At almost every site, nobody could answer. The hours had been released. Their destination was unrecoverable.
Read that precisely, because the obvious reading is wrong: this is a finding about evidence, not about outcome. The report concludes that little robust evidence remains — not that the program failed, and not that the time did not move. What stopped was the counting.
Six sites were studied closely. One had collected data continuously throughout. That site could answer the question, and its answer was that care had improved.
One in six, and the variable was not competence
Program, training, materials and intent were broadly equal across all six. The thing that separated the site with an answer from the five without one was that somebody there kept performing a small unrewarding task for years after anyone was still asking about it.
The other five are not a cautionary tale about discipline. Released time is genuinely difficult to observe — it appears as a nurse having a few more minutes on a round and never as a line in a report. Capturing it means deciding in advance what to capture, folding collection into the daily routine, and sustaining that long past the point where it is interesting. Every one of those competes with the work itself, and none of them fails loudly when it slips.
What to do with this in the week the tool arrives
The useful part is not the warning. It is that one site managed it, and what they did was ordinary: identify the number that would move if this worked, begin collecting it before the change lands, and keep collecting after the attention moves on.
Two things make that survive. The number has to be small enough that collecting it never competes seriously with the work, and it has to have a named owner with a review date in the calendar — not a team, and not an intention. That was enough to be the only people in the room who knew.
We redesign how the work gets done so the counting outlives the launch window: Decide what you will count, and who will still be counting it next spring →
Keep reading: Five hours a week came back · AI Transformation
This is the thinking behind a Trusted AI Culture.
Source: Sarre S, Maben J, Griffiths P, Chable R, Robert G, "The 10-year impact of a ward-level quality improvement intervention in acute hospitals: a multiple methods study," NIHR Journals Library, Health Services and Delivery Research No. 7.28, August 2019. The program was "Productive Ward: Releasing Time to Care," introduced in English NHS hospitals from 2007.
Written with AI in the loop: my idea, AI drafted and sharpened, my judgment on the way out. Every word is mine to stand behind. — Darren