Choosing AI for EHCP work? Three things a demonstration won't show you

If you lead a SEND service, there's a good chance you've already sat through a demonstration of an AI tool for EHCPs. With 718,800 plans in England and fewer than half of new plans issued within 20 weeks (DfE, June 2026), the appeal of producing a draft in minutes is obvious.

We’re having more conversations with councils across England that are weighing up different AI tools for EHCP work, and it's clear how difficult it can be to tell the difference from a demonstration alone.

The tools on offer vary considerably. Some are general AI models with a SEND-themed front end - they can read advice reports and write text in an EHCP format. Others have been designed specifically around EHCPs, the statutory framework, andwhat makes a quality plan.

In a short demonstration, both will produce a fluent, professional-looking draft. Here we have highlighted three things that the demonstration is unlikely to show you - and that are worth testing before you buy.

1. Whether the time saved remains saved

Drafting time is the easiest thing to measure, so it's usually where suppliers start. If a first draft used to take a day and now takes an hour, that's a real gain.

The catch is that a fast draft which is vague or inaccurate doesn't remove work - it moves it. Someone still has to notice that an outcome can't be measured, or that a need in section B has no matching provision in section F. If that's picked up at QA, it costs a senior practitioner's time. If it's picked up after the plan is issued, it can become a complaint, a mediation, or one of the 3,700 tribunals related to EHC plan content in 2025.

There's also the risk of inconsistency. Fluent text is persuasive, and a plan that reads well is easier to approve without close checking. Over time, quality can drift down while every individual draft looks perfectly acceptable.

So ask suppliers about the whole journey of a plan, not just the drafting step - and about the quality of plans that were issued, not only the speed of first drafts.

2. Whether the tool knows what a good plan looks like

Modern AI is very good at reading long documents and writing well-structured text. Give a general model some advice reports and an EHCP template and it will fill in sections A to K convincingly.

What it won't bring by default is an understanding of what makes a good EHCP. A strong plan needs a clear Golden Thread between needs, outcomes, and provision. Provision needs to be specific and quantified, so it's clear what should be delivered, and outcomes need to be specific enough to measure progress. Without that knowledge built in, a tool matches the shape of an EHCP rather than its purpose.

You can usually see the difference in a single example:

• Too vague: Access to speech and language support as needed.

• Clearer and more specific: A speech and language therapy programme set by a qualified SaLT, reviewed each term, and delivered by a trained teaching assistant for 20 minutes, four times a week.

This is the kind of distinction an AI tool designed around EHCP quality should recognise. It shouldn't simply improve the wording - it should identify why provision is unclear or insufficiently specified and help the practitioner address it.

A useful test is to use your own plans. Take a few anonymised EHCPs your QA team considers weak and see what each tool identifies. Does it find the same quality issues your experienced reviewers would - or simply produce a better-sounding version?

3. What happens when the advice disagrees

Advice rarely arrives as a tidy, consistent set. An educational psychologist recommends small-group support while the school describes a need for one-to-one help. A speech and language assessment is two years old. Health advice hasn't arrived. Spotting and resolving these things is a core part of a case officer's skill.

A general model asked to draft from conflicting advice may smooth the conflict over - choosing one version, or blending both into wording that sounds reasonable. The practitioner may never know there was a disagreement to resolve.

For example, one professional report might say a child can work independently for around 20 minutes, while another says they need continuous adult prompting to remain engaged. A general AI tool could combine those into something plausible such as “X can work independently with appropriate adult support.”

It reads well, but the contradiction has disappeared. A tool designed around EHCP quality should instead highlight the conflicting evidence and ask the practitioner to resolve it, rather than making that judgement for them.

A well-designed tool does the opposite. It surfaces gaps and contradictions so the practitioner can decide, and it shows where every line of the draft came from. That evidence traceability lets practitioners check AI-generated content quickly, see which source information it came from, and identify anything that isn't supported by the evidence. It also means that when a family or a tribunal asks why something is in the plan, the answer is there.

It's worth asking a supplier to show you this live, using a case where you already know the advice conflicts.

There's more to weigh up

These three points are a starting place rather than a full evaluation. Councils weighing up AI for EHCP work will also want to think about:

• How a tool reflects the statutory framework and SEND Code of Practice

• How it applies their own local standards

• Whether it applies the same quality checks to every plan across the service

• How practitioners stay in control of decisions

• How children's data is protected  

• How they'll know, a year on, whether any of it made a difference

Most of all, it helps to be clear about the problem you're trying to solve before you speak to suppliers. Clearing an annual review backlog, improving consistency between teams, and strengthening plans that are being challenged are different problems, and they don't all point to the same tool.