Most school AI roadmaps are not wrong about the dates. They are wrong about the order.
A leadership team sits down in October and blocks out three terms: training in the autumn, classroom pilots in the spring, policy review in the summer. It reads well in a development plan. Then spring arrives and the pilots have nowhere to land, because nobody owns the decision about which tool a teacher may use on student work, and that question has been sitting in an inbox since September.
The dates were fine. The sequence was impossible.
What should a school's AI roadmap actually contain?
A school's AI roadmap should contain five gates rather than five terms, and each gate should carry an owner, the evidence that would show it has been passed, and the date the leadership team will check.
I call this the Five-Gate Roadmap, and I offer it as my suggested way of thinking rather than a tested method:
The Five-Gate Roadmap is the way I suggest leadership teams sequence a year of AI work: as five gates rather than five terms. Each gate is a condition that has to be true before the next one can deliver anything, so the roadmap records, for each gate, who owns it, what evidence would show it has been passed, and the date the leadership team will check. The five gates are ownership, position, permission, practice and permanence. A plan that schedules a later gate before an earlier one is true is a calendar, not a roadmap.
A calendar tells you what month it is. A roadmap tells you what has to become true before the next piece of work can produce anything, which means it can also tell you that you are standing in the wrong phase. That is the one thing a plan can do that you could not have worked out on your own.
Why do most school AI roadmaps stall?
They stall because the ground under the second phase was never laid, and the published evidence suggests that is the ordinary condition rather than the unlucky one.
Start with permission, because it is the gate schools skip most often. Gallup and the Walton Family Foundation found that just 18 percent of US K-12 teachers receive formal guidance on how AI tools should be used at work, 48 percent receive only informal guidance, and 34 percent receive none at all. That survey reached 2,069 public school teachers between February 9 and March 2, 2026, and it was published on May 26, 2026. A year into near-universal adoption, four teachers in five are working from hearsay.
Now put that beside what leaders believe about their own progress. Microsoft's 2026 AI in Education Report, published on June 24, 2026 and based on a PSB Insights survey of 3,345 respondents across six countries, reports that 58 percent of education leaders say their institutions are already implementing or scaling AI, while 53 percent of educators have not had formal AI training.
Those two findings describe one school from opposite ends of the building. Leadership believes it is scaling. Staff are self-teaching and guessing at the rules. In England the gap is starker: the Tony Blair Institute for Global Change's report Generation Ready: Scaling Safe, High-Quality AI in England's Schools, published on January 22, 2026, reports that 91 percent of teachers using AI are entirely self-taught, and that only 8 percent of school leaders have made any organizational change to address generative AI.
That second figure is the argument of this article in one number. Almost every school has AI in it. Almost none has changed how it works. The activity is real and the sequence has not started, and what gets a roadmap past this is not more ambition. It is refusing to schedule gate four while gate one is still open.
Gate one: who owns this, and how does a decision get made?
Gate one is passed when a named person owns AI decisions and there is a route by which a question becomes an answer in under a fortnight.
Not a working group. Not a title with nobody in it. A person, and a route. The test I use is deliberately mundane: a teaching assistant finds a tool she wants to use with a named child next week. Who does she ask, and how long until she has an answer? If the honest answer is "she would email the head and hope," gate one is open, however many people sit on the AI committee.
Evidence it has been passed. A question asked by someone junior, answered in writing, inside two weeks, by the person whose name is on the decision. One case is enough. The route either works or it does not.
What happens if you skip it. Every later gate turns into a request nobody can grant. This is why so many pilots die quietly: not from opposition, but from an unanswered question. Who should own AI strategy and who belongs in the room when those decisions get made go further, because the composition of that group decides how fast the route runs. If this gate is not closed by half term, it is the only thing on your roadmap that matters.
Gate two: where do we actually stand?
Gate two is passed when the leadership team can state its position on AI in a form a governor or a board member could challenge.
Not a feeling. Not a tool count. A position: what staff are already doing, what the school has decided, what it has not, and where the gaps sit. The reason this gate comes second rather than first is that you cannot baseline a school that has no owner, because nobody has the standing to ask staff what they are really doing and be told the truth.
The scale of the honest answer usually surprises people. RAND's survey panels found that 45 percent of principals reported having school or district AI policies as of spring 2025, while 54 percent of students and 53 percent of English language arts, math and science teachers said they used AI for school. That report was published on September 30, 2025, and the distance between those two numbers is the gap most leadership teams are standing in without having measured it.
Evidence it has been passed. A written position, one or two pages, that someone outside the senior team has read and argued with. The seven dimensions of AI readiness is the diagnostic I use for this, and what benchmarking actually tells you is worth reading first if the question in the room is whether you are behind.
What happens if you skip it. You buy the solution to somebody else's problem, because a roadmap built on an unmeasured starting point prioritizes whatever the last conference talked about. This is a half-term job, not a year-long audit, and a plan that puts its readiness review in the summer has mistaken the diagnosis for the treatment.
Gate three: what are staff allowed to do?
Gate three is passed when a member of staff can say, without checking, which tools they may use, on what information, and what they must never put into one.
This is the Gallup finding made local. Eighteen percent formal guidance is not a training problem. It is a permission problem, and training before permission produces the worst outcome available: staff who now know how to do something nobody has told them they may do. They either stop, which wastes the training, or they proceed privately, which is how shadow AI starts.
Permission has three parts and most schools write only the first. What is approved. What is forbidden. And what to do when something is neither, which is the part that decides whether the first two survive contact with a new product. What a school AI policy should contain and what to require before approving a tool take that further, and the question I am asked most often in sessions, whether staff may put student data into ChatGPT, belongs in writing at this gate rather than in a corridor conversation.
Evidence it has been passed. Ask three members of staff at random, separately, what they are allowed to use for a task involving a named student. If you get three answers, the gate is open.
What happens if you skip it. Adoption happens anyway, unevenly and invisibly, and the policy you eventually write has to undo habits instead of shaping them. Gates two and three can run together, and they are the only pair that comfortably can.
Gate four: is one job being done differently by people who did not volunteer?
Gate four is passed when a single, named task is routinely done a new way by staff who are not enthusiasts.
The qualifier carries the whole gate. Enthusiasts prove nothing about an organization, because they would have found the tool with or without you. What tells you something is a reluctant colleague doing a specific job differently in March than she did in September, and not having to think about it.
Name the job narrowly. "Drafting first-pass written feedback on Year 10 essays, which the teacher then edits" is a job. "Using AI to save teachers time" is a hope. One job done properly beats five pilots, which is the pattern behind why AI pilots so often go nowhere and the ceiling that AI champions programs hit when the next step needs a decision rather than a demonstration.
Evidence it has been passed. You can name the job, name two people doing it who did not ask to, and say what they stopped doing to make room.
What happens if you skip it. You reach the end of the year with training delivered, policy written and nothing visibly different, which is the outcome most likely to persuade a governing board that the whole thing was overstated. This gate needs a full term, and moving from experimentation to organization-wide adoption is its mechanics.
Gate five: will this survive the people who started it?
Gate five is passed when the new way of working would continue if the person who introduced it left at Christmas, and you can show someone else what it changed.
Two halves, both easy to leave out. The first is capability: the practice has to sit in a routine, a template or an induction rather than in one person's head. RAND's American School District Panel, published on April 8, 2025, found district AI training rose from 23 percent to 48 percent of districts between fall 2023 and fall 2024, with leaders prioritizing teachers' concerns and fears over instructional uses. It describes a sector that has got much better at starting and has said far less about what holds. Training is an event; capability is what is left a term later, which is the distinction behind building capability rather than dependence and what should happen after the initial training.
The second half is evidence, and it buys next year's permission. You do not need a research design. You need to say what the school does differently and what you watched to know, which is the ground proving AI is having an impact covers.
Evidence it has been passed. The practice appears in a document a new member of staff would be given, and one page exists that a governor could read. This is the final term's work, and it is what turns a year of activity into a position you can build on.
The Mistake I See Most Often
The mistake I see most often is a leadership team working hard at gate four while gate one is still open, and reading the lack of progress as resistance.
It looks like energy. There is a champions group, a tool everyone likes, a training day in the diary. What there is not is anyone who can say yes, so every decision the group needs goes somewhere else and comes back slowly or not at all. Six months later the conclusion in the room is that staff were not ready. Staff were ready. The route was never built.
I taught, then became an assistant headteacher, then Director of Digital Strategy in a further education college, which is where I learned this the expensive way: the plans that worked were never the ambitious ones. They were the ones that asked, before each phase, what has to be true for this to work, and then went and made that true first. Across the leadership teams I work with now, the most useful hour is rarely spent planning the year. It is spent finding the earliest gate still open, which is almost never the one the plan says they are working on.
That diagnostic use of the five gates is worth more than the sequencing. Walk them in order and stop at the first gate you cannot produce evidence for. Whatever it is, it is this term's work, whatever the roadmap says.
The one-page roadmap: what to put in front of your governors
A roadmap worth presenting is one page: five rows, five columns, and no prose.
Copy this into a document or into five rows of a spreadsheet. The columns are the point: a roadmap with dates but no evidence column is a wish list, because nothing in it can ever be shown to have been passed.
| Gate | The question it answers | Owner | Evidence it has been passed | Check date |
|---|---|---|---|---|
| 1. Ownership | Who decides, and how fast? | A junior colleague's question answered in writing inside two weeks | ||
| 2. Position | Where do we actually stand? | A two-page written position someone outside the senior team has argued with | ||
| 3. Permission | What may staff do, with what? | Three staff, asked separately, give the same answer | ||
| 4. Practice | What is being done differently? | One named job, two non-volunteers doing it, one thing they stopped | ||
| 5. Permanence | Will it survive us? | The practice is in an induction document, and one page exists for governors |
Three rules make the difference between a page that works and a page that decorates a shared drive.
One owner per row, and a person rather than a role. Evidence written before the work starts, not after, because evidence invented at the review is always met. And a check date for each gate, separate from the gate before it, so that a stall is visible in November rather than in July.
When you take it to a board, take it with the gates marked as passed or open. A governing body that can see you are at gate two in February will ask much better questions than one shown a timeline with everything on track. What to report to governors about AI sets out the rest of that paper.
When should a school run two gates at once?
A school should run two gates at once when the later one needs a long lead time it cannot recover, and only after naming what it will do if the earlier gate fails.
The objection to strict sequencing is fair. A year has three terms, five gates is more than three, and some work has to start early: a training day booked a term ahead, infrastructure that takes months. A rigid queue can turn a planning discipline into an excuse for a slow year.
So here is the honest limit of the framework. Gates two and three overlap well, because a baseline and a permission statement inform each other. Gate five's capability half can be designed while gate four runs, and should be. What does not work is gate four before gate three, or any gate before gate one. Those are not preferences about order. They are dependencies, and the later work does not deliver without them.
If you do run ahead, write the fallback in the same row: what happens to that booked training day if permission is not settled by then. A plan that names its own failure mode is sturdier than one assuming the earlier gate will come good.
What should leaders do this term?
Take the plan you already have, put the five gates beside it, find the earliest gate for which you cannot produce the evidence, and stop working on anything later than it.
Not forever and not as a reset: just for this term, and visibly, so staff can see the sequence rather than guess at it. Most leadership teams I work with find their open gate is one or three, which means the work in front of them is smaller and less interesting than the work in the plan. It is also the only work that makes the plan possible. Then write the year as five rows with five owners and five check dates, and take it to your governing board as gates rather than terms. The understatement will do more for your credibility than a timeline ever has.
The next step
If your leadership team is sequencing a year of AI work and wants the order challenged before it reaches a governing board, this is the kind of work I support through AI strategy sessions for schools. If the open gate turns out to be your own confidence rather than the plan, coaching for school leaders is the closer fit. If you would rather read than book anything, I write a newsletter for school and trust leaders on the decisions this work turns on.
Sources and further reading
- Gallup and the Walton Family Foundation, "Most Teachers Receive No Formal Guidance on AI Use," May 26, 2026. Survey of 2,069 US K-12 public school teachers, February 9 to March 2, 2026. news.gallup.com
- Microsoft, "2026 AI in Education Report," Microsoft Source, June 24, 2026. PSB Insights survey of 3,345 respondents across six countries. news.microsoft.com
- Tony Blair Institute for Global Change, "Generation Ready: Scaling Safe, High-Quality AI in England's Schools," January 22, 2026. institute.global
- RAND, "AI Use in Schools Is Quickly Increasing but Guidance Lags Behind: Findings from the RAND Survey Panels," September 30, 2025. rand.org
- RAND, "More Districts Are Training Teachers on Artificial Intelligence: Findings from the American School District Panel," April 8, 2025. rand.org
Dan Fitzpatrick is the founder of The AI Educator, a Forbes contributor and the author of bestselling books on AI in education. Read more about Dan.
