Draft the plan, but keep the child in the room.

There is a quiet contradiction sitting in the guidance schools are now reading, and it is worth pulling into the light.

Updated Department for Education guidance, developed with the Chartered College of Teaching and Chiltern Learning Trust and downloaded more than forty-four thousand times, tells schools they may use AI to draft a pupil's SEND support plan. Draft a staff guide for a particular need. Generate a template letter home. In the same document, it tells staff to be extremely cautious about inputting a child's information, forbids AI from writing a statutory Education, Health and Care Plan, and forbids it from deciding whether a child has a need at all.

Read those two instructions together and you have the whole design problem of AI in a school, in miniature.

Because here is what drafting a support plan actually involves. You are handling the most sensitive information a school holds: a named child, the difficulties they have, the help they get, the things that go wrong on a bad day. That is exactly the data you least want to send anywhere. And most AI tools work by sending it somewhere. The words you type travel to a server you do not own, in a place you cannot see, governed by terms that can change while you sleep. “Be extremely cautious” is sound advice, but it puts the whole weight of safety on a tired teacher at half past nine on a Sunday, deciding word by word what is safe to paste.

I taught for twenty years before I started building AI. I know that teacher. I have been that teacher. And I know that a warning is not a safeguard. A warning is a way of moving the risk onto the person with the least time to carry it.

There is a cleaner answer, and it is not “use AI less”. It is “keep the work where the data already lives”.

When the model runs on the device in front of you, the difficult calculation disappears. The child's information does not travel, because there is nowhere for it to travel to. It is drafted, checked and saved on the laptop the school already owns, and it never crosses a border, a network or a company boundary. The caution the guidance asks for stops being a discipline you have to maintain and becomes a property of how the thing is built.

This matters more, not less, because of where SEND policy is heading. The Schools White Paper published in February proposes recording far more day-to-day provision in Individual Support Plans, with statutory EHCPs kept for the most complex needs. Whatever you make of the reform, the practical effect is more plans, drafted by more classroom teachers and SENCOs, for more children. The drafting burden is about to grow. If every one of those plans is a small decision about whether a child's data is safe to type, we have built a system that leaks by design and asks exhausted people to plug the gaps.

The technology to do this properly is no longer the hard part. Meta's on-device research team reported this year that billion-parameter models now run in real time on an ordinary phone, and that four-bit compression holds most of a model's quality at a quarter of the memory. A well-trained small model can now rival one many times its size. The same techniques that put a capable model on a handset put one on the mid-range laptops schools already have. The old objection, that a local model cannot be good enough, is ageing fast.

So the question in front of schools is not really “should we use AI for SEND”. The guidance has answered that: carefully, yes. The real question is where the drafting happens, and who has to worry about it when it does.

At Aevum we build for the second answer. Tools designed so a child's information stays in the room it is about, and never comes to us. Built accessibility-first, because the founder spent two decades watching who gets left out when technology is designed for the average child. Trained in the UK, on UK infrastructure, accountable to UK law, and honest about how it is made.

Draft the plan. Get the evening back. Just do not make a teacher trade a child's privacy for the time. Keep the child in the room.