How the AI drafting works, and where it stops

What the AI does and does not do

The drafting assistant produces a first draft of the school’s written contribution to an education, health and care (EHC) plan, laid out in the twelve sections set out in regulation 12 of the SEND Regulations 2014 (A to K). It works from the evidence the school already holds: pasted or uploaded professional reports, staff feedback forms and parent contributions already on the pupil’s record, and the pupil information the school has entered.

It does not:

• Write the EHC plan. The local authority does that. This is the school’s advice, which the authority may draw on.
• Decide anything: not whether to request an assessment, not what a pupil needs, not what provision is required.
• Send anything to a local authority, a parent or anyone else. Drafts stay in the school’s account until a member of staff exports them.
• Run without the school switching it on for each pupil.
• Use any other part of the product. Deadlines, the communication log, safeguarding alerts, targets and forms involve no AI.

Consent first

Drafting is blocked for every pupil until a member of school staff records on the pupil’s record that the school has a lawful basis and consent for AI processing. The record states, in the app, what will be pseudonymised and what cannot be. The school is the data controller and decides whether to use the feature at all; many schools use everything else in Sendiphany and never switch it on.

Pseudonymisation before anything leaves the system

Before any text is sent to the AI provider, the pupil’s first name, last name and full name, their date of birth in the common ways it is written in UK reports, the school’s name, and the names of the key worker, local authority caseworker, parent, feedback-form teachers and the requesting member of staff are replaced with placeholder tokens such as [CHILD_FIRSTNAME] and [PERSON_1]. The model is told to use those tokens and never to guess a real name. When the draft comes back, the real values are re-inserted inside Sendiphany and any token the model invented is neutralised. The substitution map never leaves the system.

Grounded in your evidence, with the gaps flagged

The model is instructed to use only the evidence supplied, to make every need, outcome and item of provision traceable to something in that evidence, and never to invent assessments, scores, diagnoses, dates, professionals or events. Where evidence is missing for a section it must say so in that section and list the gap under a “missing information” list returned with the draft. Sections that are weak or need professional judgement are listed as “quality flags”. The evidence blocks are marked as quoted third-party material so that any instruction hidden inside a pasted report is ignored. Outcomes are asked for in specific, measurable terms and provision in quantified terms, matching paragraphs 9.66 and 9.69 of the SEND Code of Practice.

The SENCO reviews and owns every word

The draft is saved as a version on the pupil’s record, together with the model used and a reference to the evidence it drew on. The SENCO (or SENDCo) edits every section in the app; nothing is marked final, compliant or approved by the software, and the system prompt forbids the model from describing its output that way. Named versions can be kept before changes. Regenerating a draft never overwrites edited text without an explicit confirmation, and the previous state is snapshotted first. Exports are Word documents the school then owns and submits under its own name. Our terms make human review a condition of use.

The provider and your data

Drafts are generated by Anthropic’s Claude models through Anthropic’s API, which is listed on our sub-processors page. Prompts and outputs are not used to train AI models. Each request records the model, token counts, duration and success or failure in our usage log for cost and reliability monitoring; that log contains no pupil content. Requests are rate-limited per user and per pupil, and the feature is unavailable once a school’s trial has ended without a plan.

DfE guidance on generative AI

The Department for Education’s guidance on generative AI in education (2025, updated 2026) accepts AI help with first drafts of support plans but is clear that statutory EHC plan content must not be written by AI without significant professional review. We have designed to that line rather than around it: the tool drafts the school’s contribution, not the plan; it works only from the school’s evidence; it flags what it does not know; and a qualified member of staff reviews and edits every word before anything leaves the school. The DfE’s product safety expectations for generative AI also ask suppliers to say which providers process data, whether prompts are retained, and that outputs require human review. This page is our answer; we have described the position cautiously and will revise it as the guidance changes.

Limits and known failure modes

• The model can misread evidence, over-generalise from a single observation, or produce fluent text that reads more confidently than the evidence supports. The quality flags help, but they are not a guarantee.
• Very long evidence is cut short; the app tells you when that happens and asks you to trim to the relevant sections.
• The model may occasionally decline to draft from material it judges unrelated to the pupil; the app explains this and the SENCO can draft manually.
• The draft reflects the quality of the evidence. Thin evidence produces a thin draft with a long list of missing information, which is the correct outcome.
• Pseudonymisation does not catch identifiers the school has not recorded (see above).

Questions

A data protection officer who needs this in a form suitable for a data protection impact assessment, or anyone who wants to ask how a specific case was handled, can write to privacy@sendiphany.co.uk. Our security page covers the rest of the system.