How ARC uses AI

Where AI appears in ARC, what is stripped out before anything reaches a model, where the model runs, and what is written down about every call.

Anyone 6 min read Updated 20 October 2026

ARC uses AI in three narrow places, all of them summarising records that already exist. Nothing decides anything, nothing is generated about a young person from outside their own record, and the output is always labelled as a draft.

This page describes what the software does. If AI is switched off for your system — it is off unless deliberately enabled — you will see "AI summaries aren't enabled on this system." instead.

Where you will meet it

On an incident. The incident page carries an AI summary panel, badged Advisory draft, holding a short paragraph drawn from the incident's own narrative and the updates recorded against it.

On a journal day. The Daily Summary section has a Draft with AI button. It writes an advisory summary of that day's logs into the Daily Summary box for you to read, correct and save — or not.

When a manager decides about Regulation 40. Ticking This incident may need to be notified to Ofsted in Manager Review drafts a Summary for Ofsted from the incident, naming children by their initials only, for the manager to copy into Ofsted's own form. ARC submits nothing. If AI is off, or the model fails, the summary is built from the record by a fixed template with no external call. It is stored on the incident and can be redrafted. See Record a Regulation 40 notification.

The journal one deserves care. The draft is not saved by ARC; but the box it lands in is part of the journal, so whatever you leave there and save becomes the record, in your name. Read it before you save it. The helper text on the page says the same thing: it is a summary "for you to review and edit".

Identifying details are stripped first

Nothing reaches a model until it has been through a pseudonymisation layer, in two passes.

The deterministic pass removes what ARC already knows for certain — the young person's full name, first name, last name, preferred name and public id, and the social worker's name, email and phone. It then sweeps the text for structured personal data: email addresses, NHS numbers (checked against their check digit, so ordinary ten-digit numbers are not mistaken for them), phone numbers, postcodes and dates. Each is replaced with a bracketed placeholder such as [CHILD_NAME_1].

The detection pass then reads the already-stripped text and points at what is left — other people's names, schools, hospitals, organisations, place names. It runs after the first pass, so it never sees the young person's own name.

That second pass only points. ARC does the replacing itself, so a model is never trusted to rewrite the text — only to read it.

Placeholders are put back only when the result is displayed, for the person who asked.

Where the model runs

Inference runs on Amazon Bedrock in London (eu-west-2), in region. This is a deliberate constraint rather than a default: cross-region inference profiles are forbidden in ARC's configuration because the EU profile routes traffic EU-wide, which would break UK-only residency. The endpoint URL itself pins the region.

Pseudonymisation is risk reduction, not an exemption

It has to be said plainly, because it is easy to over-read. Stripping identifiers substantially reduces what leaves ARC — it does not make the text non-personal data, and it is not a GDPR exemption, because the text may remain re-identifiable. It is paired with the residency constraint above, not offered instead of it.

Every call is written down

Each pass writes one row to ARC's AI audit log: the home, the young person, the person who asked, the purpose, which record it came from, which detector and model were used, how many items were replaced, and the categories of those items.

It records counts and categories, never the text and never the personal data that was stripped.

What is advisory, and what gets saved

This is the distinction to hold on to.

Journal summary Incident summary
Generated Only when you select Draft with AI Automatically, when the incident is saved and each time an update is added
Saved by ARC No Yes — stored on the incident and shown when the page loads
Where it lands In the Daily Summary box, for you to edit In the AI summary panel

The incident summary is regenerated in the background, so the page waits for it rather than blocking the save; if it fails, it fails quietly and the record is untouched. Every summary panel ends with the same line: "Advisory only — verify against the record before acting."


Was this helpful?

Tell us what was missing — help@arccare.co.uk. We reply within one working day, Monday to Friday.