Twelve questions to ask before you sign anything.

The single most useful question to ask an AI supplier is where the data goes, because almost every other risk follows from the answer. After that, the questions worth asking are about accountability, exit, and evidence. The Department for Education now publishes standards defining what a generative AI product must meet to be considered safe in an education setting, which gives a school something firmer to point at than a feeling.

Why is a school at a disadvantage here?

Because a bank buying AI has a procurement department whose job is to say no, and a primary school has a business manager with four other jobs. A bank writes data residency into the contract because it has the leverage to insist. A school gets whatever the vendor ships by default, which is generally whatever was cheapest to build.

That is the asymmetry these questions are meant to reduce. None of them requires technical knowledge. All of them are uncomfortable to answer evasively, which is the point.

The twelve questions

• Where is our data processed, physically, and under which country's law? Ask for a place, not a policy.
• Is any pupil work or staff input used to train your models, ever, including in aggregate or anonymised form?
• What happens to our data if we stop paying, and how do we get it out? Ask for the format.
• Can a teacher see why the system produced a particular answer? If not, how do we defend a decision to a parent?
• Which of your claims have been independently evaluated, and by whom? Ask for the protocol, not the headline.
• Which DfE product safety standards do you meet, and which do you not? A supplier who claims all of them without qualification has not read them.
• Who is accountable when the output is wrong, and what does your contract say about that?
• What is your policy on children's data specifically, under the ICO's Children's code?
• What happens to our workflow if your company is acquired, changes terms, or closes?
• Has the tool been tested with pupils who have SEND, and what did that testing find?
• What is the total cost across three years, including the training time and the staff hours to adopt it?
• What does the tool refuse to do, and why? A supplier who cannot name a limit has not thought about one.

How should a school weigh cost against benefit?

By counting the hidden costs, which are rarely the licence. Training time, the staff hours lost while everyone adapts, the cost of migrating off it later, and the cost of the thing going wrong once. A tool that is free at the point of use is a bill you have not been handed yet, payable in dependency and in data.

The return side is harder, and honest suppliers will say so. Time returned is measurable if you bother to measure it before and after. Attainment effects are not measurable in a year by a school on its own, and anyone offering you a figure for them is offering you something they cannot support.

The questions we would rather you did not ask us

Ask us for benchmark scores and we will tell you they are unpublished until our pre-registered evaluation protocols are complete, which is true and also inconvenient. Ask us for a launch date and we will not give you one. Ask us whether our safeguarding layer is ready for use with children and the answer is no, it is research, and we have published why.

We would rather lose a comparison on the honest answer than win one on a number we cannot defend. Hold every supplier to that, including us.