Put one question to the heads of twelve schools in the same trust: who decides whether a new AI tool may be used in your school? You will not get one answer. You will get the trust, the head, the IT lead, and a pause. The pause is the most honest of the four, because it belongs to the person who has realized, mid-sentence, that a tool is already running in three classrooms and nobody can name who let it in.
That is not a discipline problem. It is a delegation problem, and it is the single most common structural fault I find in trusts and districts that otherwise lead AI well.
Who owns AI in a multi-academy trust?
The trust owns the decisions that have to be the same in every school. Each school owns the decisions that only change what happens inside it. And the trust board owns the fact that both kinds are being made at all, whether it has written that down or not.
That last sentence is the one that catches people, so it is worth being precise about where it comes from. A multi-academy trust is a single legal body running several schools, and the Academy trust handbook 2026, which takes effect on 1 October 2026, puts it plainly: the board has "collective accountability and responsibility for the academy trust," and it is the board that determines "what, if any, governance functions are delegated to the local tier." Delegation is a decision the board makes, in writing, in a scheme it must approve and should review each year.
Read that alongside a second fact. The handbook does not mention artificial intelligence once. It does not need to. Accountability did not change when AI arrived; only the number of decisions did. A board that has never discussed where AI decisions sit has not avoided owning them. It has delegated them by accident, which is the one form of delegation nobody can audit.
Why the question gets answered twice, in opposite directions
The evidence that trusts are answering this question inconsistently is now on the record. In January 2026 the Education Policy Institute published Exploring multi-academy trust approaches to artificial intelligence, drawing on roundtables with fourteen trusts responsible for 442 schools between them. Its finding on structure is careful and, read closely, uncomfortable: "decisions will inevitably happen at multiple levels and school leaders and teachers will need to make decisions that best suit their context."
Inevitably is doing real work in that sentence. Decisions will land at several tiers whether or not anyone planned for it. Some trusts in the study managed tool approval centrally. Others simply added AI paragraphs to policies they already had. One leader described the gap between the two ends of the same trust: "Go to one of our schools and it's VR, AI, immersive learning. Go 300 yards down the road and there's no tech at all."
Three hundred yards. Same board, same funding agreement, same accountability, two entirely different answers to what a child is allowed to use.
The American version of this question has the same shape and different vocabulary. In March 2026 the ILO Group published Adaptive Governance for AI in K-12: Organizing for a Competency, Not a Tool, which argues that districts keep treating AI as something to procure and protect when it is closer to a capability the organization has to build and hold. It separates decisions into three tiers: strategic questions of why and whether, tactical questions of how and who, and operational questions of what. The labels differ from the English ones. The failure is identical: nobody writes down which tier holds which question, so the question is settled by whoever is standing nearest to it.
The Same Answer Test
The Same Answer Test is three questions I ask of any AI decision in a trust or a district before deciding where it belongs: If every school answered this differently, would that be a problem or just a difference? When it goes wrong in one school, whose name is on the letter that goes home? And could a single school actually carry this, with the people and the hours it has? A decision that needs the same answer everywhere, or that lands on the center when it fails, or that no single school can carry, belongs to the center. Everything else belongs to the school, and taking it anyway is not support. It is collection.
This is my suggested way of sorting the decisions, not a validated instrument, and its value is in how quickly it settles arguments that have been running for a term. Most AI questions fail none of the three and belong to the school. A small number fail all three, and those are the ones a board should be able to name without looking.
The second question is the sharpest of the three, because it is the one leaders answer honestly. Ask a trust executive who decides which AI tools teachers may use and you get a discussion. Ask who writes to the parents when a tool has done something it should not have, and the room goes quiet and then says the chief executive, or the trust. That answer has located the decision. It just has not moved it yet.
The four decisions a trust cannot delegate
Four decisions fail that test in nearly every trust and district I work with. They are not the interesting decisions. They are the ones that bind.
The data decision. What staff and students may put into an AI tool, and who completes the assessment that stands behind that permission. Under UK data protection law the duty is unambiguous: where processing is likely to result in a high risk to people's rights, the Information Commissioner's Office requires the controller to carry out a data protection impact assessment, a written appraisal of the risk, before the processing starts rather than after. Children count as vulnerable in that judgment. The controller is the body that decides why and how the data is used, which in a multi-academy trust is the trust and not one of its schools. The EPI roundtables found leaders struggling with that duty honestly: they "did not necessarily understand the tools well enough to do so," and asked for a national source of information on the products everyone is using. That is a reasonable request. It is not a reason to push the assessment down to a school that understands the tool even less.
The approval decision. Which tools may be used at all, and what would take one off the list. This is the decision a school can genuinely make for itself and should not, because twelve schools approving separately produce twelve different suppliers holding the same children's data. If you want the detail of how a single approval should work, I have written elsewhere about what a school should require before approving an AI tool; in a trust the list itself is the artifact, and it has one owner.
The answering decision. Who is named when it goes wrong, and on what date they look at it again. Not a role, a person. A trust that cannot say who is accountable when an AI tool gets it wrong has not distributed that accountability across its schools. It has mislaid it.
The floor decision. The minimum every school must meet, stated as a floor and not a ceiling. This is the one trusts get backwards most often. A floor says: every school will have told its parents something by December, and no school will use a tool that is not on the list. A ceiling says: this is what AI in our trust looks like, and it is the same everywhere. Floors travel well across schools that are genuinely different. Ceilings do not, which is why they get quietly ignored by the schools furthest from the center.
What stays with the school, and why taking it is not support
Three things belong to the school, and a central team that collects them has made itself the bottleneck for work it cannot see.
What AI is for here. The trust can say what AI is for across the trust. Only the school knows whether its own pressure this year is reading at Key Stage 2, staff turnover, or a leadership team that has changed twice. The purpose has to be written where the pressure is felt.
Who does what, and when. Sequencing, staffing, which year groups, which subjects, which term. The EPI report quotes a trust leader on exactly why this cannot be run centrally: "All of our academies are different. They're on different curriculums, they're on different timetables, they're on different exam boards."
What this school tells its own parents. The trust can supply the words. It cannot supply the relationship, and a letter that reads as though it came from a head office will be read that way.
There is a test for whether a central team has overreached. Look at how long a school waits for an answer. If a head has been waiting three weeks for permission to let a department try something that affects nobody outside the building, the decision is in the wrong place, however tidy it looks on the scheme of delegation.
The Mistake I See Most Often
Trusts centralize the enthusiasm and delegate the liability. It should be the other way around.
I see it in the same order almost every time. The central team runs the AI strategy day, builds the training, picks the platform, and writes the vision. All of that is visible, energizing work, and it is genuinely useful. Then a head asks whether their staff can put student names into a tool, and the answer comes back that schools should make that judgment in line with their own data protection responsibilities. At that moment the trust has kept the part that makes it look like a leader and handed away the part that makes it one.
I have sat on the receiving end of this as a trustee, and the question that exposes it is not a technical one. It is simply: which of these decisions would come to this board if it went wrong? If the honest answer is all of them, and the board has never seen any of them, the scheme of delegation is describing a trust that does not exist.
But we already have one AI policy for the whole trust
This is the most reasonable objection, and it is half right. A single trust-wide policy is a genuine achievement and most trusts do not have one. It answers the second question, though, not the first. It says what the rules are. It does not say who may change them, who may make an exception, and who finds out when a school has quietly stopped following them.
The EPI study caught the gap precisely. One trust had mandated AI literacy training across all its schools, and its own leaders acknowledged that "this is not happening in practice." The policy was trust-wide. The decision to act on it was never located anywhere, so it stayed where it had always been, with whoever had time. That is what an unowned decision looks like from the outside: a document everybody agrees with and nobody is behind.
A policy is a set of answers. Ownership is knowing who is allowed to change them. If you are working out which of the two you actually have, the distinction between an AI strategy and an AI policy is the place to start.
What to do before your next board meeting
Four moves, and none of them takes a working group.
- Write down the last five AI decisions made anywhere in the trust, with the date each was made and the name of the person who made it. Include the ones made by not deciding. This takes about twenty minutes and is usually the most revealing document a trust produces all year.
- Run each one through the Same Answer Test. Mark it center or school. Expect to find at least one decision sitting in the wrong place and at least one sitting in no place at all.
- Check the scheme of delegation against that list. The board must approve that scheme and should review it annually, and the 2026 handbook takes effect on 1 October. If AI decisions appear nowhere in it, that is the paragraph to draft now, while the review is open, rather than next September.
- Give every decision you have located a review date and an owner. A decision with neither is a preference. The people who need to know which it is are the heads who have to act on it on Monday.
If you want a single question to open the board discussion with, it is the one that settles who the AI group in your trust is actually for: who should sit on the group that makes these decisions, and what are they allowed to settle without leaving the room.
Sources and further reading
- Department for Education, Academy trust handbook 2026: effective from 1 October 2026, GOV.UK.
- Lily Wielar and Jon Andrews, Exploring multi-academy trust approaches to artificial intelligence, Education Policy Institute, funded by the Nuffield Foundation, January 2026.
- Information Commissioner's Office, When do we need to do a DPIA?, UK GDPR guidance and resources, last updated 13 October 2025.
- ILO Group, Adaptive Governance for AI in K-12: Organizing for a Competency, Not a Tool, March 2026.
Where this usually goes next
Most trusts that work through this find the sorting is the easy half. The hard half is that several of the decisions they have just located centrally have never actually been made, by anyone, and the board is a week away from asking about them. That is strategy work rather than policy work, and it is the kind of thing I support through AI strategy sessions with trust and school leadership teams, where the output is a short list of decisions with names and dates against them rather than another document. If you would rather read first, I write about this each week in the newsletter, and the broader question of how to create an AI strategy people actually use is where this argument came from.
Dan Fitzpatrick is the founder of The AI Educator, a school trustee, and an adviser on AI in education to government bodies including the UK Department for Education, KHDA in Dubai and the Ministry of Education in Kazakhstan. More about Dan.
