The AI summary card sits at the top of the right-hand column of an incident, badged Advisory draft. It is a reading aid for a record that has grown long: a narrative, a dozen updates, linked logs, debriefs.

It is not the record, and it is not a decision.

What it does

ARC generates a short summary — two to five sentences of plain prose — from everything shown on the incident: the overview, location, staff and police involvement, the Regulation 40 status, the missing episode, each physical intervention and its record, linked events, return home interviews, notifications, follow-up tasks, the manager's review and note, and the updates recorded on it.

Debriefs are left out of the summary saved on the incident, because everyone who can open the incident sees that summary and debriefs are restricted. A summary asked for by someone allowed to read a debrief includes it, is shown to that person only, and is not saved.

On an incident this happens for you: the summary is generated in the background when the incident is created, and refreshed each time an update is added. You don't ask for it.

The card carries the provenance underneath: "Generated [date] from this incident's report. Advisory only — verify against the record before acting."

What it will not do

It is instructed to use only what the record states — not to infer, speculate, embellish or offer advice. So a summary that says nothing about risk is telling you the record says nothing about risk. It is not telling you there isn't any.

Read the record. The summary exists so that reading it is faster, not so that it can be skipped.

What is sent, and where

ARC assembles the text server-side from the record itself — the page only names which incident. That text then goes through the pseudonymisation layer before it reaches any model:

The young person's recorded identifiers are replaced with placeholders, along with structured personal details. So are the names of staff recorded on the incident and of the other young people in the home.
A second pass looks for incidental personal information written into free text that the first pass couldn't know about.
The placeholders are swapped back only when the summary is shown to you.

Inference runs on Amazon Bedrock in London, in-region only — the strict UK-only residency rule that applies to every AI feature in ARC.

Every generation is audited: who ran it, against which record, and how much was stripped. The audit never holds the text or the personal information itself.

It is saved to the incident

Unlike the summary on a journal day, an incident's summary is stored on the incident and appears on the incident PDF under the heading AI Summary (advisory), with the same warning to verify against the record.

That makes it visible to anyone who can read the incident, including anyone you share the PDF with. Two things follow:

It is not private working-out. Treat it as part of what you are handing over.
It still isn't the account. What you wrote is the account; the summary is a restatement of it, generated by a machine and labelled as such.

If nothing happens

If the card says "AI summaries aren't enabled on this system.", the feature is switched off for your ARC. Nothing is missing from the incident itself — the summary is chrome over a record that is complete without it.