Basic AI training is no longer enough. Two sets of data published this autumn point the same way: employers are asking for fewer people who can simply use AI, and teachers use it least in the tasks where a wrong answer would land on a child. Train staff for the point where their use stops, because that is where the judgment is.
Most AI training answers one question: how do I use this? That question has a short life. The one that replaces it is harder to teach and matters more: when should I trust what comes back, and when should I not use the tool at all?
Are employers still asking for basic AI skills?
Not in the way they were. Workday's Global Workforce Report, published on October 5, 2026, analyzed job requisitions from roughly 550 employers using Workday's recruiting software between September 2025 and July 2026. Demand for basic AI skills, such as simple prompting, rose through late 2025, peaked in January 2026 and then fell 25 percent. Demand for hands-on skills, such as building AI tools and automating workflows, climbed 51 percent over the whole period. The report puts it this way: "Companies have moved on from wanting people who can just use AI to wanting people who can build with it."
Hold that at its true size. It is a software vendor's release, drawn from job advertisements at its own customers, and an advertisement records what an employer writes down, not what the work needs. A school is not hiring AI engineers either. What transfers is the direction. Once everyone can do the basic thing, naming it stops separating anyone.
The same report carries the number I would put in front of a leadership team. In a September 2026 survey of 6,000 full-time employees, 79 percent said they know what skills they need to succeed, and 66 percent said their employer helps them develop those skills. That is a 13-point gap between knowing what is needed and being helped to get it.
Where does staff AI use actually stop?
It stops where a wrong answer is hardest to catch and most likely to land on a person. Teacher Tapp, the daily survey of teachers in England, shows it. In an article for the Chartered College of Teaching's Impact journal on how teachers are using AI in 2026, published on May 18, 2026, Teacher Tapp's Karen Wespieser reported answers from between 8,000 and 10,000 teachers surveyed in November and December 2025, reweighted to match the Department for Education's workforce census. Among teachers who had used AI for schoolwork, 54 percent had used it for lesson planning, 50 percent for quizzes and 41 percent for letters to parents. Only 11 percent had used it for lesson delivery.
A second Teacher Tapp comparison, reported by EdTech Innovation Hub on October 15, 2026, sets what teachers predicted in April 2023 (5,254 teachers) against what 8,837 teachers reported in September 2026. Use beat the predictions for planning and parent communication: about 56 percent of primary senior leaders now use AI for parent communications, against about 28 percent predicted. It fell short for the work that touches individual pupils. Primary teachers' use of AI to analyze attainment data was about 7 percent, against 34 percent expected, and secondary teachers' use of AI for marking in the previous month was about 7 percent, against 31 percent.
Those two marking figures are not the same measure. The first article asks whether a teacher has ever tried AI for marking, and 46 percent of AI users had, though Wespieser notes that the use is more limited in depth. The second asks about the last month. Taken together they are consistent with teachers who tried it and did not keep going, though two surveys that ask different questions cannot prove it.
The American picture has the same grain. The EdWeek Research Center's surveys show the share of US teachers with no AI training falling from almost 60 percent in October 2024 to 42 percent in winter 2025-26, with 9 percent receiving ongoing training. EdWeek's reporting on that data found that schools often start with basic uses such as planning lessons and writing emails to parents, and that moving from AI as an efficiency tool to teaching and learning has been harder.
Is this a motivation problem or a judgment problem?
It is a judgment problem. Look at why the minority who have never used AI say they have not. In Wespieser's data, 56 percent of them cited the reliability of the information, 46 percent cheating and plagiarism, and 46 percent pupil overreliance. Thirty-five percent said it does not save time, because the output has to be checked. Only 19 percent said they lacked the confidence to learn how to use it.
Read that last number twice, because it is the smallest on the list. The leading barrier is not "I cannot work out the tool." It is "I cannot be sure of the answer, and checking it takes as long as doing it myself." That is a judgment problem, and a course on prompting does not touch it. This is my reading of the data, and the sample is the small group of non-users, so I hold it loosely. But it fits what the training shelf looks like. Most of what is on it was built for the question the Workday data says the market has stopped asking.
The Judgment Line: how do you find where your own staff stop?
Ask them, and put the answers in order. I call the result the Judgment Line, and I offer it as a way of looking, not a tested method.
The Judgment Line is the point on a list of your staff's tasks, ordered by what a wrong answer would do to a person if nobody caught it, where AI use stops. Training that stays below the line teaches people to use the tool. Training that reaches the line teaches them to check it, and to say when it should not be used.
Building one takes an afternoon.
- List the tasks. Ask a dozen staff from different roles for the twelve to fifteen things they do in a typical week that AI could touch. Use their words, not yours.
- Order them by consequence. Put the tasks in a column from "a wrong answer gets fixed before anyone sees it" to "a wrong answer shapes what happens to a named person." The question for each is who is affected if the error is never caught.
- Mark the use. Beside each task, mark whether staff use AI for it commonly, sometimes, rarely or never. A two-minute anonymous poll is enough.
- Draw the line. Find where "commonly" gives way to "rarely." That is your line, and it is rarely where leaders guess.
- Take the first task above it. Not all of them. The one nearest the line is the subject of your next session.
A worked example
This one is hypothetical, a secondary school invented for illustration.
| Task | If wrong and uncaught | Staff use of AI |
|---|---|---|
| Draft a newsletter paragraph | An editor reads it before it goes out | Commonly |
| Write quiz questions for next week | The teacher uses them within days and notices | Commonly |
| Draft a letter to parents about a trip | It reaches families under the school's name | Sometimes |
| Draft pupil report comments | A parent reads it as the teacher's view of their child | Rarely |
| Give first-pass feedback on essays | It shapes what a pupil believes about their own work | Rarely |
| Summarize attendance patterns for leaders | It shapes what leaders do about named pupils | Never |
In this school the line falls between the parent letter and the report comments. The next training session is therefore not about prompts. It is about report comments: what an invented observation looks like, who reads the draft before it goes home, and what the teacher checks it against. My piece on whether teachers should use AI to write school reports goes through that one task in detail.
What I See in Practice
The first training day works because it teaches something with an immediate payoff, and the second one often does not, because the payoff has moved. I have trained more than 150,000 educators across more than 30 countries. Across the leadership teams I work with, the pattern after the first session is consistent. The questions that come up stop being about the tool. They become: how do I know this is right, and what happens to me if it is not? Those are the right questions, and a session built around features has nothing to say to them.
The usual leadership response is another session on the same ground, with a better tool or a newer feature. That is what happens after the initial AI training when nobody has asked what the first session left unsolved. The staff who stopped at the line did not fail to learn. They reached the edge of what the training covered.
When is staying below the line the right answer?
Often. A teacher who does not use AI to mark essays, or a deputy who will not paste attendance data into a chatbot, may be exercising exactly the judgment you want. The aim is not to push everyone above the line. It is to make sure that every task above it gets an answer on purpose rather than by default.
For each task above the line, there are three honest answers. Leave it alone and say why. Allow it with a named check: who reads the output, against what, and what counts as wrong. Or reserve it for a person, which is what which decisions AI should never make in your school is about. What you cannot do is leave the question unanswered and let the gap between "allowed" and "used" fill with private guesses.
What should the next training session contain?
The session below the line teaches a tool. The session at the line teaches a check. In practice that means four things:
- Real output with real mistakes. Use drafts from your own staff's work, with their permission, including the confident ones that were wrong.
- One check for one task. Teach what to compare the output against and what a fault looks like, for the single task nearest the line.
- A stated boundary. Say which tasks are reserved for a person and which need a named reader, in a sentence staff can repeat.
- A return. Ask each person to bring one real output to the next session, so the check is practiced on live work, not only seen once.
Good looks like this in a building: staff can tell you, task by task, whether they use AI and why, and nobody is guessing whether they are allowed. It is a different thing from a high usage number, and measuring AI maturity by something other than usage explains why that number can mislead.
If your leadership team is working out what staff need once the first round of training is done, this is the kind of work I support through AI training for schools and leadership sessions. For the wider question of which human skills matter more as the basic ones become universal, start with what skills matter when everyone has access to AI.
Sources and further reading
- Workday, Global Workforce Report: AI Is Rewriting Jobs More Than It's Cutting Them, October 5, 2026. Job-requisition data from about 550 employers, September 2025 to July 2026; employee survey, September 2026. A vendor publication.
- Karen Wespieser (Teacher Tapp), From novelty to normal: How teachers are using AI in 2026, Chartered College of Teaching, Impact, May 18, 2026. English teachers, November and December 2025.
- EdTech Innovation Hub, AI use in schools exceeds teacher predictions for lesson planning but lags for marking and data, October 15, 2026, reporting Teacher Tapp data (5,254 teachers in April 2023; 8,837 in September 2026).
- EdWeek, More Schools Are Providing AI Training for Teachers. Is It Any Good?, May 2026, reporting EdWeek Research Center surveys of US teachers.
Dan Fitzpatrick is the founder of The AI Educator and writes and speaks on how leaders should respond to AI. More about Dan.
