Two inventories decide what a school can prove. Neither was ever anybody's job.

Two inventories now decide whether a school can evidence its SEND provision, and neither was ever anybody's job. One is what was put in place for a child, when, and what happened next. The other is which AI tools are actually in use in the building, and what each one was allowed to read. Several pieces of guidance landed this month that assume a school holds both. Most schools hold neither, and that is not a failure of diligence.

What changed this month

Four things arrived close enough together that the pattern is easier to see than any one of them alone.

• Keeping Children Safe in Education 2026 came into force on 1 September. It is the first edition to deal with AI directly, and it brings generative AI inside filtering and monitoring.
• The Department for Education updated the filtering and monitoring core standard on 16 September. It says provision should be reviewed when new technology is introduced, not only once a year, and names AI-generated content as something the system has to cope with.
• The academy trust handbook was updated on 17 September. It is the document setting out what a trust must do with public money, and trust boards are asked to be assured about risk.
• Ofsted's September inspection changes require a school to show that pupils with barriers to learning are achieving well. Show, not assert.

Read plainly, the filtering standard means that every time a member of staff starts using a new AI tool, the standard has already been triggered. And a board cannot be assured about a risk it has no list of.

Why is the first inventory so hard to produce?

Because provision is delivered continuously and recorded occasionally, usually by the person with the least time. In the schools I worked in, the record of what a child actually received lived in three or four places at once: a provision map someone updated when they could, a teacher's own notes, an intervention log kept by a teaching assistant, and a memory.

None of that is negligence. It is what happens when recording is a separate act from doing, performed afterwards, by someone already behind.

The scale is the part that does not get stated often enough. Department for Education figures published on 11 June, counted at the January 2026 census, put over 1.8 million pupils in England with identified special educational needs, 20.8 per cent of all pupils. Of those, 538,547 hold an education, health and care plan and 1,319,780 are on SEN support. So for every child with a plan, roughly two and a half have recognised needs and no statutory document at all. That second group is where the evidence problem concentrates, because nothing external forces the record into existence.

Why is the second inventory harder still?

Because nobody adopted these tools through procurement. A teaching assistant found something that drafts a support plan in two minutes. A head of year found something that rewrites a letter home. Neither was being reckless; both were doing a job with no guidance attached.

So the list a board is now expected to hold is a list of decisions nobody recorded making. And the question that matters about each entry is not what the tool does. It is what it was allowed to read.

The Information Commissioner's Office audited 28 edtech providers whose products are used across primary and secondary schools, publishing on 24 June. It made 596 recommendations and providers accepted 98 per cent of them. Not one of the failings was exotic: providers unclear whether they were processor or controller, contracts that were not detailed enough, incomplete data flow mapping, out-of-date privacy information. If the suppliers were unclear about their own arrangements, a school holding no list at all is not an outlier.

What happens when the record is missing but the duty is not

The school does the work and cannot show it. That is a different failure from not doing the work, and it is punished the same way.

Analysis reported by Tes on 11 September found that schools with the highest proportion of pupils with SEND are five times more likely to miss the Ofsted inclusion standard. Some of that is genuinely about provision and capacity. Some of it is about whether a school can show what it did, for whom, and with what effect, inside the window an inspector is looking at. Those are different problems and they need different answers.

The deadline makes it concrete. Schools receiving the inclusive mainstream fund have to publish an inclusion strategy on their website by 31 December 2026. Schools Week reported in March that the fund is worth around £400 million a year over three years, roughly £48,000 for a secondary school and about £14,000 for a primary in 2026 to 2027. The strategy is meant to set out the commonly occurring needs in your cohort and what you ordinarily provide. That is the first inventory, written down, with ten teaching weeks to do it.

The detail I keep going back to

FE Week assembled forty local SEND reform plans from council websites and from drafts requested from officials, publishing on 10 September. Seventeen of the forty say next to nothing about post-16 provision. Thirty-six flag staffing as a risk to delivery. Five councils acknowledged their post-16 arrangements were not yet embedded, and none of the five listed that on a risk register.

The problem was known. It was written into the plan, in prose. It was not recorded anywhere that would make anyone act on it.

That is the whole argument in one example. Knowing a thing and holding the record of it are not the same, and only one of them survives contact with an inspection, an audit or a board meeting.

What we are building because of it

Aevum Teacher Manager is being built on the assumption that the record should be a by-product of doing the work, not a separate act performed afterwards by someone who is already tired. Marking produces the evidence of what a pupil struggled with. The paperwork module drafts reports and letters from that marking and from the records already in the building, rather than from a blank page.

On the second inventory, our answer is narrower and more structural: if the tool runs on the school's own hardware, the question of what it was allowed to read has one answer, and it is checkable. That is not a claim that local inference is privacy. A paper by Jonghyun Chung and Sanket Badhe published on 8 June argues exactly the opposite, and they are right: running locally answers only where the computation happened, not what the system could read or what authority it held. Those have to be designed and shown separately.

All of it is in development. Nothing here is a finished product, no school is yet using it, and there are no outcome figures because none have been measured. When there are, they will come with the protocol that produced them.

Own the model. Keep the data. Inspect the reasoning.