The training day went well. People stayed behind to ask questions, the feedback forms came back warm, and within a month a good share of staff were using AI most weeks. Then the leadership team went looking for the thing the training was supposed to produce, and it was not there. Same hours. Same meetings. Same pile on a Sunday evening.
That gap is now measurable. Bett's Working It Out report, published on August 27, 2026 and based on a YouGov survey of 1,033 UK teachers, found weekly AI use had doubled in a year, from 26% to 52%. In the same survey only 35% said they were working fewer hours and 55% said they were working the same. Enthusiasm is cooling with it: the share expecting to use AI more next year fell from 68% to 47%.
The short answer: nothing happens, unless a leader changes the work
Nothing happens after the initial AI training unless someone changes what the work is. Training moves a skill into a person. It cannot move a decision into an organization, and the decision is the part that was missing. The instinct at this point is to book a second training day, and a second training day is usually the first one again, with better catering.
The saving is real. It just never adds up
Both stories are true at different altitudes: the minutes come off the task, and the week does not change.
The best measurement we have of the saving comes from the Education Endowment Foundation's Teacher Choices trial on using ChatGPT for lesson preparation, published on December 12, 2024 and independently evaluated by the National Foundation for Educational Research. Across 259 teachers in 68 English secondary schools preparing Key Stage 3 science lessons, which in England means students aged 11 to 14, teachers using ChatGPT spent 56.2 minutes a week on preparation against 81.5 minutes for the comparison group. That is a saving of 25.3 minutes a week, close to a third of the task. An expert panel reviewed the resources without knowing which had been made with ChatGPT and found no difference in quality. The share of teachers saying they spent too much time on lesson preparation fell from 49% to 26%.
So the minutes are real, and they are real in the measured, blinded, independently evaluated sense that most claims about AI and workload are not.
Now look one level up. Gallup surveyed 23,717 employed US adults between February 4 and 19, 2026 and found that 65% said AI had improved their productivity and efficiency, while only around 10% strongly agreed that AI had transformed how work gets done. The same shape appears outside education, in organizations that have change teams and none of a school's timetable constraints.
Twenty-five minutes came off one task. Nobody decided where those twenty-five minutes should go, so the week decided instead, and a week always decides in favor of whatever was already overdue. That is the whole mechanism. It is not a failure of the tool, and it is not a failure of the teacher.
The distinction worth holding on to: a task-level saving becomes an organization-level change only when somebody makes a decision in between. Absent that decision, the efficiency is absorbed, and it is absorbed so quietly that nobody can name the day it disappeared.
Why more training is the most popular wrong answer
More training is the popular answer because it fits in the calendar, not because the evidence points at it.
The supply problem is close to solved. RAND's American School District Panel survey, reported by Melissa Kay Diliberti, Robin J. Lake and Steven R. Weiner on April 8, 2025, found that the share of US districts providing teachers with generative AI training doubled in a year, from 23% to roughly half, with another 26% planning to start. The same study found most of that training was voluntary, aimed at teachers' concerns, confusion and fears rather than classroom application, and that 11 of the 14 district leaders interviewed had built it themselves because they could not find qualified outside expertise.
Since then the coverage has grown and the depth has not. Education Week's reporting in May 2026 on its own Research Center surveys found the share of teachers reporting no AI training at all fell from 60% in October 2024 to 50% in fall 2025 to 42% in winter 2026. In that same winter 2026 reading, 22% had been to more than one session and 9% had anything ongoing. Jessica Garner, a managing director at the educator association ISTE+ASCD, put the problem in one line: "We can't stop with efficiencies."
Read those two findings together and the position is clear. Introductions have roughly doubled. Ongoing support sits below one in ten. The sector has become good at the first hour and has barely started on the second year. That reading is my interpretation of the two surveys, not a finding either one makes.
The first day cannot do the job on its own, and the reason is structural rather than a matter of quality. A first training day teaches what I call linear innovation, doing what we already do, faster. That is the correct place to start and the honest thing to sell. But a faster version of an unchanged week is still an unchanged week. The second move, asking whether the work should have that shape at all, is not a training outcome. It is a leadership decision, and no trainer in the room has the authority to make it.
The Release Question
The Release Question is three questions I ask before booking any further AI training: What will people stop doing when this works? Who is allowed to decide that it stops? And where will the time that comes back show up, in a name, a date or a number? Training that survives all three changes the work. Training that fails any of them changes a person and leaves the work exactly where it was.
Take them one at a time, because each one fails differently.
What will people stop doing when this works? Most leadership teams can answer what staff will start doing, in detail, with examples. Asked what stops, they reach for words like "streamline". Streamlining is not stopping. Stopping has a name attached to it, and somebody notices.
Who is allowed to decide that it stops? This is where the answer usually goes quiet. A teacher who has halved the time a task takes cannot unilaterally decide the task no longer needs doing, or that the template it feeds into can be retired, or that the meeting it was prepared for can be shorter. Those are permissions, and permissions live above the person who found the saving.
Where will the time show up, in a name, a date or a number? If the answer is "people will have more time", the time has no destination and will not arrive. If the answer is "Year 8 reports move from a Sunday to a Wednesday planning period, and the deputy head checks in November whether that held", the time has somewhere to go.
I offer this as a filter I find useful, not as a validated instrument. Its value is that it fails loudly. A leadership team that cannot answer all three in a single meeting has found the reason the last training day did not land, and it is not the trainer.
What has to happen after the first training day
Three things have to happen, in this order: a subtraction, a destination, and a second attempt.
A subtraction. Name one thing that stops, in writing, with the date it stops. A form, a template, a standing agenda item, a report nobody reads. This is the same test I set out in moving from experimentation to organization-wide adoption, where the blunt version applies: if nothing stopped, nothing was adopted. It was added.
A destination. Say where the released capacity goes before it is released, and say it in the terms the organization already uses. Planning time, a reduced deadline, a smaller meeting, an earlier finish, a task moved off a weekend. Capacity with no destination is not a benefit. It is a vacancy, and vacancies get filled.
A second attempt. The training day produces one attempt on a safe example. Nobody becomes competent at anything on one attempt, and the thing that makes the second attempt useful is that it happens on real work with somebody to check it. This is what the 9% figure in the Education Week data is really measuring. It is not a measure of enthusiasm for training. It is a measure of how rarely anybody gets a second go.
None of the three requires another trainer, and that is the point.
What I Have Learned From Working With Organizations
The organizations that change are not the ones that liked the training most.
Across the training I have run with schools, colleges and businesses, and the organization-wide programs among them, the pattern that holds is this: the sessions that get warm feedback and the sessions that change anything six months later are not the same sessions. What separates them is almost never the content. It is whether somebody senior had already decided what would stop.
The question that changes a room is not "what can AI do?" It is "what are you going to stop doing?" Ask it of a leadership team and the silence that follows is the actual finding. It is not awkwardness. It is the sound of a group discovering that they have been treating an efficiency question as a training question, and that the part they own is the part nobody has done.
The most common version of this I meet is a leader who is frustrated with staff for not using AI enough, in an organization where every single thing that was on the calendar last year is still on the calendar this year. Staff read that accurately. They can see that the work has not moved, which tells them the tool is an extra, and an extra is the first thing dropped in a difficult week.
When more training is the right answer
More training is right when usage is low rather than flat, and the two look nothing alike.
If weekly use sits with a handful of enthusiasts, if staff cannot say what they are allowed to put into a tool, or if people are quietly avoiding it because they are not sure whether they would be in trouble, then you do not have a work design problem. You have a permission and confidence problem, and training is part of the answer. I set out how to read that situation in how leaders build AI confidence across a workforce.
The Release Question is for the other case: broad use, steady or rising, and a set of outcome measures that have not moved. That is the case the Bett survey describes, and it is the one where a second training day is a way of looking busy.
Two honest limits on the evidence above. The EEF trial covers one subject, one key stage and one task, it was published in December 2024, and it did not measure any effect on pupils. And the tempting next step, measuring whether the saved time showed up, runs into a separate trap I have written about in why usage is a poor measure of AI maturity: a number that rises while the work stays identical is measuring activity, not progress.
What to do before the next training day
Four things, and none of them take a day.
- Put The Release Question to your leadership team in a single meeting and write down the answers, including the blank ones.
- Name one task that stops this term, with a date and a person.
- Name the destination for the capacity that releases, in your own scheduling language, and say who checks in November whether it held.
- Book the second attempt before you book the second course. A follow-up on real work, with someone to check it, beats another introduction to another tool.
Then, and only then, look at what training you still need. You will usually find it is shorter, later, and aimed at fewer people than the one you were about to book.
The first AI training day was never going to transform anything by itself, and that is not a criticism of it. It did its job. What happens next is not a training question at all. It is the oldest leadership question there is, asked about a new tool: what are we going to stop doing?
If your leadership team has done the training and is now looking at a workload picture that has not moved, this is the kind of work I support through AI training and professional development programs built around what the organization is going to change, not just what staff are going to learn. You can see how I approach that on AI training for schools and colleges.
Sources and further reading
- Cerys Turner, "Teachers' use of AI soars, but has little impact on working hours", Tes, August 27, 2026, reporting Bett's Working It Out report (YouGov survey of 1,033 UK teachers): tes.com
- Education Endowment Foundation, "Choices in EdTech: using generative AI (ChatGPT) for KS3 science lesson preparation", Teacher Choices trial, evaluated by NFER, published December 12, 2024: educationendowmentfoundation.org.uk
- Gallup, "Rising AI Adoption Spurs Workforce Changes", survey of 23,717 employed US adults, February 4 to 19, 2026: gallup.com
- Melissa Kay Diliberti, Robin J. Lake and Steven R. Weiner, "More Districts Are Training Teachers on Artificial Intelligence: Findings from the American School District Panel", RAND, April 8, 2025: rand.org
- Education Week, "More Schools Are Providing AI Training for Teachers. Is It Any Good?", May 2026, reporting EdWeek Research Center surveys: edweek.org
Dan Fitzpatrick is the founder of The AI Educator, a Forbes contributor and an international keynote speaker who has delivered AI training to more than 150,000 educators across more than 30 countries. More about Dan.
