How Ofsted looks at AI, and what you need to be able to show

Ofsted does not inspect AI tools. It inspects the decisions your service made about using one, and whether those decisions held up for the children in your care.

Its own guidance is unusually direct about that boundary:

“We do not look at AI as a stand-alone part of our inspections, and we do not directly evaluate the use of AI, nor of any AI tool. However, inspectors can consider the impact that the use of AI has on the outcomes and experiences of children and learners.”

And on what the evaluation actually looks at:

“Any evaluation of the use of AI will ask whether the provider has made sensible decisions – the same as we would with any other tool.”

Ofsted names three risks an inspector may ask about

They are data protection, safeguarding, and bias or discrimination.

“Inspectors may ask what steps the provider has taken to make sure their use of AI properly considers these risks. Were those risks considered when introducing the AI application? Do they have any assurance processes to identify emerging risks? Any evaluation that Ofsted makes is about the provider’s decision-making, what they have considered, and the impacts on children and learners, not about the tool itself.”

The guidance says inspectors may ask. It sets out no list of evidence a provider must produce. What it describes is a conversation about your reasoning, and that conversation is easier when the reasoning was written down at the time it was made.

Ofsted’s own examples include case notes and meeting summaries

Two sentences in the guidance describe almost exactly what a lot of services are already doing.

Ofsted names “support staff using it to help take case notes” as an example of AI use in this sector. Later, in a list of examples, it gives this one:

“if a local authority uses AI-generated summaries of child protection conferences, inspectors may want to understand how it makes sure these are accurate”

The question in that example is how you know the summary is right. It applies to any tool that touches a record, whether or not there is a model inside it.

The consultation closed on 28 September and did not mention AI

Ofsted’s consultation, “Improving the way Ofsted inspects children’s social care”, closed on Monday 28 September 2026. Reading the document, we could not find any mention of AI, technology or record-keeping. On the evidence in front of us, it does not change how Ofsted looks at AI. When the response is published we will read it and say here whether that still holds.

It did propose two other changes. One is a new evaluation area for independent fostering agencies, covering the quality of foster carer recruitment, assessment and support. The other is a move from “best fit” to “secure fit” grading, where a provider would show how it meets the full set of criteria rather than the strongest few.

There is related work already published: “Artificial Intelligence (AI) in case recording”, written by Research in Practice for the National Workload Action Group and published by the Department for Education in September 2025. Useful reading, with two caveats. It cites Ofsted’s earlier AI guidance, which has since been withdrawn and replaced by the page quoted above. And its own recommendation is that national guidance on AI in case recording should be developed, so none exists yet.

Where the model runs is a choice Ofsted leaves to you

Cloud AI tools are open to you under this guidance. Councils up and down the country already run them, with published data protection impact assessments. Ofsted’s question is whether choosing a tool was sensible, and whether you can show your working.

Practitioners have given their own reasons to take that question seriously. The Ada Lovelace Institute’s report “Scribe and prejudice?” (February 2026) interviewed 39 social workers across 17 local authorities. Some of them reported AI transcription producing content that was never said: references to suicidal ideation a client had not expressed, and words inserted into the record of what a child had told them. The British Association of Social Workers issued a position statement on the use of Palantir Technologies in social work services on 4 June 2026, arguing that social work must remain a human profession centred on care, relationships, dignity and social justice.

These are reports from practitioners rather than proven findings. They make the case for the evidence Ofsted describes: a trail showing what a tool was asked to do, what it produced, and who checked it before it became part of a child’s record.

Three questions about any tool that writes into a record

A record written with a tool’s help becomes part of a child’s file, and one day that child may read it. Ofsted’s questions come down to whether you can account for how it got there.

• Why this tool was introduced, and what was weighed at the time.
• What happens when it gets something wrong, and how anyone would know.
• Where the record of what it did actually lives, and who can see it.

In Chrysalis, what the model writes is a draft until a person reads it and signs a declaration adopting it, and every action lands in a tamper-evident audit log. Founding places start at £1,000 a month and £2,250 in year two, sized to the service. Apply for a founding place, or book a twenty-minute walkthrough and see the real screens first.

Also in Guides: What Ofsted asks a fostering agency for when it gives notice of inspection.