I have been working in schools for long enough to have written A LOT of inclusion plans. I know exactly what happens to most of them.
They get written in a hurry, usually at the end of a long day, usually because a deadline is approaching. The strategies are the same ones I have always used — not because I have reviewed the evidence for each child's specific profile, but because they are the ones I know. The ones that feel safe. The ones that have worked before, sometimes, for some children.
This is not negligence. It is what happens when one person is responsible for 80 pupils, three statutory timelines, a parent meeting on Wednesday, and a supply teacher who needs briefing by 8am.
When AI-assisted planning tools appeared, the conversation focused almost entirely on speed. How quickly can the AI draft a plan? Can it save the SENDCo an hour? Can it reduce the paperwork burden?
These are real questions. But they are the wrong first question.
The right first question is: where do the strategies come from?
The problem with “AI-generated” strategies
An AI that generates strategies from a thin or opaque evidence base is not evidence-based planning. It is autocomplete with a professional veneer.
I have seen SEND planning tools that cite “research” without naming a single source. I have seen AI outputs that suggest strategies that contradict NICE guidance for the specific need type being addressed. I have seen plans that recommend interventions with no published evidence of effectiveness for SEND populations — a problem identified in Pegram et al (2022).
The AI is not the problem. The evidence base behind the AI is the problem.
If a strategy fails — if a child does not make progress, if a review cycle shows no impact — the SENDCo needs to be able to explain why that strategy was chosen. “The AI suggested it” is not a defensible professional answer. It is not what the SEND Code of Practice means when it requires provision to be evidence-based. It is not what Ofsted means when inspectors ask for evidence of impact.
Professional accountability requires a traceable chain between the child's documented profile, the research that supports the chosen approach, and the outcome that was expected. That chain has to exist before the plan is written, not assembled retrospectively when something goes wrong.
What an evidence base actually needs to do
For AI-assisted planning to genuinely support professional practice rather than just accelerate it, the evidence base behind the AI needs to do several specific things.
What we built
OMNIA's Research and Strategy Library now holds 160 named independent evidence sources. All of them are fully structured and AI-citable. All of them are jurisdiction-tagged — so a strategy drawn from WWfSEND ASD research is marked as globally applicable, while a reference to the DfE Inclusive Mainstream Fund guidance is shown only to English schools, because it is English legislation and showing it to schools in other contexts would be confusing and misleading.
The library includes systematic reviews, NICE clinical guidelines, DfE statutory guidance, specialist charity research, inspection frameworks, and academic research syntheses. It includes sources specifically focused on SEND populations — not just general school improvement evidence applied to SEND by inference.
When OMNIA generates a plan for a child with dyslexia, the AI draws on the assessment scores, cross-assessment flags, the pupil's own voice, and the Available Interventions Register for that school — and it cites the specific evidence source and strand that informed each strategy recommendation. The SENDCo can see, in the plan itself, that a Precision Teaching recommendation came from the EEF Teaching and Learning Toolkit's metacognition strand, with a note on effect size and evidence strength.
The library is live-scanned. When a source is updated — when EEF revises a strand rating, when a new NICE guideline is published — the scanner flags the change for review before it affects any plan output. A human reviews and approves every evidence update before it goes live. Because the AI drafts and the SENDCo decides — and that principle has to extend to the evidence base itself, not just the plan content.
Why this took longer than the features
Building the plan generator took weeks. Building 141 structured, jurisdiction-tagged, AI-citable evidence sources took longer.
It would have been faster to hard-code a list of eight well-known sources into the AI prompt and ship. Many tools do exactly that. It produces plans that look evidence-based and cite recognisable names, and most people reviewing them would not notice the difference.
But a SENDCo who has to defend a provision decision to a parent, an Ofsted inspector, or a tribunal deserves more than the appearance of evidence. They deserve the real thing — a traceable, named, need-specific, jurisdiction-appropriate evidence trail that they can follow from the child's profile to the strategy to the source.
That is what we built. It is slower to build than a feature. It is harder to explain in a demo than a button that generates a plan in 20 seconds. But it is the thing that makes the plan worth generating.
OMNIA's Research and Strategy Library is available to browse at omnia-inclusion.com/evidence. Founding school places are open until July 2027 — five places globally.