AI support plan writing: optional, labelled and always human-reviewed

AI can help busy staff turn notes into a first draft. It can also write confident nonsense about a vulnerable person. Inclusion Studio treats AI as an optional drafting helper, clearly labelled and always checked by a person. This page explains the rules and what the prototype shows.

Can AI help with learner support plan writing?

Writing up a review takes time. Staff may have a learner’s quote, a plan and a delivery record, and need to pull them into a short summary and a suggested next step. A draft can save that first step. It can also help a tutor who finds writing hard, as long as they check what it says.

The risk is that a draft reads as fact. That is why the rules below exist. Our guide to using AI safely in learner support sets out the wider good practice.

How do AI drafts work, step by step?

In the prototype, using a fixed sample. In production, the same flow would apply if a provider switches AI on.

1. A person asks for a draft. Nothing is drafted automatically.
2. The draft uses only the record in front of the person. For a review, that is the learner’s quote, the agreed support and the delivery record.
3. The draft is labelled. It says it is a sample or AI-generated draft and must be checked.
4. The person edits. They change, delete or replace any of it.
5. The person decides. Only when they save does anything become part of the record. The record shows the text was reviewed by a person.
6. The learner involvement is recorded separately. A draft is never a substitute for talking to the learner.

What will AI never do?

These are design rules, not aspirations:

• Diagnose a condition or infer one.
• Refuse, reduce or approve support.
• Make decisions about a learner.
• Share information outside the organisation.
• Invent delivery evidence.
• Be trained on case records.
• Hold a therapeutic conversation with a learner, or act as an open-ended chat for learners.
• Read or label a learner’s emotions.

The selected provider is Gemini on Google Cloud Vertex AI. Its location for real learner data, retention and terms will be documented and assessed before real learner data reaches it. There will be no silent fallback to a different provider. See AI principles.

Who sees what?

Drafts follow the same permissions as the underlying record. A draft can use only information the requesting person is allowed to see, enforced on the server. A tutor cannot use AI to read a learner’s records they could not otherwise access. See permissions and roles.

What is in the prototype today, and what is being built?

In the prototype (fictional learners only):

• A “Use sample AI draft” button in the review form. It inserts a fixed text built from the fictional learner’s quote, delivery record and next step.
• A disclosure: “Sample draft from the demo record. Check and edit before saving.”
• An “Evidence behind the draft” view listing the three sources and a statement that no learner information is sent to an AI service.
• An “AI draft available, review before saving” strip on the learner summary.

The draft is deterministic text, not output from a model.

Being built now:

• A provider-level switch to turn AI off or on.
• A synthetic or mock adapter first, so AI behaviour can be tested without real data.
• Cost controls and a kill switch.

Later, only after it passes evaluation: permission-scoped review drafts with source references. AI features can stay switched off during the pilot, and each one is enabled only after accuracy, safety and location checks show it does better than the same task without AI.

Provider: Gemini on Google Cloud Vertex AI is the selected AI provider. It has been exercised with synthetic data only, in a preview environment, at an EU multi-region endpoint accepted for preview use. Where real learner data would be processed is not decided, the live application uses a mock, and no real learner data has been sent to any AI provider.

Not decided: where real learner data would be processed, and whether AI is used with real data at all. That is a documented decision to be made, with the Data Protection Officers of pilot providers involved, before any real record is processed.

Is this safe for special category data?

Support records often contain health information, which is special category data under UK GDPR. Sending that to an AI service needs a lawful basis, a Data Protection Impact Assessment and a contract with the provider. See our guides to special category data in education and DPIA for learner support software. Inclusion Studio does not make this decision for you.

Who is this for?

Tutors and assessors who want to save time on write-ups. Inclusion leads who want clear rules before staff use AI. Data protection officers who need to know what is processed where. See data protection officers.

What does this look like in practice? (fictional example)

A tutor is preparing a review for a fictional learner. They select the sample draft button. The text it inserts restates the learner’s quote, notes that four sessions were recorded and suggests a short check-in after each session, adding that the change should be checked with the learner first. The tutor deletes the second sentence because it does not match what the learner actually said, adds their own observation and saves. The record shows the learner’s involvement as “To discuss with learner”, because the conversation has not happened yet. The draft helped with the first sentence only, and the person made every decision. This is a fictional example.

Related guides and next steps

• Using AI safely in learner support
• Reviews and AI principles
• Join the pilot.