Four people asked for the same record this week

One trust answered 133 subject access requests last year. Three years earlier it answered 16.

Tes reported on 22 October 2026 that Dixons Academies Trust had seen a 731 per cent rise. Outwood Grange Academies Trust went from 14 requests in 2018-19 to 376, spending around £100,000 a year of staff time on them. The piece names generative AI as the main driver of both volume and complexity, with some requests clearly AI-drafted and citing the wrong legislation.

A subject access request is not hard because somebody asked. It is hard because everything about one child sits in eight different systems, and a person has to go and find all of it by hand inside a statutory deadline. The asking got cheap. The finding did not.

It was one of four times this fortnight that somebody wanted a school's record of a child and could not read it.

The other three

On 15 October the Department for Education opened a call for evidence on school exclusion reform, closing on 19 November. It asks what schoolwork a child actually gets during the first five days of a suspension or permanent exclusion. Those are usually the days when nothing is written down. The paperwork that exists is about the exclusion, not the learning. Anyone who has tried to reconstruct a suspension eighteen months later, for a tribunal or an annual review, knows that gap. The evidence is absent rather than wrong, and absent is much harder to argue with.

On 20 October Schools Week reported the National Governance Association's survey of more than 3,000 governors and trustees. Sixty-eight per cent now put SEND support in their top three budget pressures; staff costs came second, on 65 per cent. It is the first time since the survey began in 2022 that SEND has overtaken staff pay. Staff pay is the largest line in any school budget, so the worry is not about size. It is about not being able to predict it. You can forecast a pay award. You cannot forecast how many children will arrive in September needing support nobody has funded yet.

And on 15 October the department published key stage 4 performance for 2025/26. Nineteen point seven per cent of pupils in state-funded schools in England had identified special educational needs, up from 18.7 per cent the year before and 14.2 per cent in 2018/19. Nearly one in five sitting GCSEs with an identified need.

Where the week disagrees with itself

Those last two do not point the same way.

The key stage 4 trend partly reflects schools getting better at spotting need, and partly a system in which help is gated behind a label. Read that way, the answer is fewer labels. The governors' survey points the other direction: schools need earlier and better information about the children already in the building, which sounds like more identification, sooner.

Neither is solved by labelling more children. What each asks for is a record of need, and that is not the same document as a record of diagnosis. One is written by the school, continuously, about what it is already doing. The other arrives from a clinic, late, and tells the school something it mostly knew.

The fourth asker was not a person

FE Week carried a piece on 15 October about an assessment tool built at a Hampshire further education college, designed to argue with the marker rather than replace them. It quotes Charlie Munger on incentives: show me the incentive and I will show you the outcome. The incentive in most AI marking is speed, and speed converges the marking towards the answer the model expects. The pupil who answered sideways loses, and disproportionately that pupil has SEND. A model reading a child's work is reading a record, and it will only read the one it was given.

What we are working on

Four parties wanted the same thing: a parent inside a statutory deadline, a government consultation, a governing board forecasting three years, and a marking model. None of them got it, because the thing they wanted mostly lives in people. The teacher who knows which lesson it goes wrong in. The SENCO who holds the pattern and has no hour spare to write it down.

That is the problem Aevum Foundry AI is working on. The honest version is less impressive than the pitch version.

We are not building a compliance product that produces evidence on demand. The binding constraint is the evening, not the database. Any system that asks a SENCO to type the record in order to have the record has already failed, because that hour does not exist.

So the approach is to make the writing-down a by-product of work already happening, on hardware the school owns, so a child's words never leave the building to become searchable. On-device is not a feature here. It is the reason a school can afford to put real detail in. A record you would not send to a third party is a record you will not write.

We have not built that. There is no product to look at and no date to give, only a problem statement we think is the right one and an architecture chosen because of it.

If a request landed at your school this morning, how many places would you have to look?