Summarising long referral documents
Referral packs run to dozens of pages. AI can pull the key facts out for whoever is deciding — as long as nobody treats the summary as a substitute for reading the pack.
Our rating
Try with care
Can help, but needs the right tool, a data protection check and a person reviewing every output.
What it looks like
Before
A 34-page referral pack: hospital discharge summary, social work assessment, medication list, two letters from the family, and a physiotherapy report.
With AI
One page under fixed headings — needs, risks, medication, who else is involved, what is missing — with page references back to the source for each point.
An invented example.
A referral pack for one person can run to thirty or forty pages: a discharge summary, an assessment, a medication list, correspondence. Somebody has to read all of it and decide whether the service can meet the person’s needs, usually today.
This is one of the better uses of AI in care, and it has one failure mode that matters more than all the others.
How it works
You give the tool the pack and a fixed set of headings. It returns a structured summary — needs, risks, medication, who else is involved.
The single most useful thing you can add is a heading for what is missing. “No GP details. No information about night-time support. Medication list is dated three weeks before discharge.” That is often more valuable than the summary itself, because gaps are what cause problems on day one.
Where it helps
Triage speed. Knowing in two minutes whether a package is plausibly within your capacity, before committing an hour to reading.
Consistency. Every pack summarised under the same headings makes comparison possible.
Catching gaps. Tools are good at noticing a heading with nothing under it, which is exactly the kind of thing a tired human skims past.
The failure mode that matters
The summary reads complete, and it is not.
A pack mentions, once, on page 30, that the person had a fall three weeks ago. It appears in a physiotherapy note, in passing. It does not appear in the discharge summary or the assessment.
A summary produced from that pack will very likely not mention it, because it appeared once, in an unremarkable sentence, in a document mostly about something else. The summary is not wrong. It is a reasonable summary. And the most important fact in the pack is not in it.
This is not a fixable flaw. It is what summarising does: it keeps the representative and drops the outlier. In care referrals the outlier is routinely the point.
So the rule is: the person who decides reads the full pack. The summary tells them what to expect and where to look. It does not replace the reading, and any process that lets it is unsafe.
Two smaller failure modes
Over-certainty. Source documents in health and social care are full of hedging — “query dementia”, “reports occasional dizziness”, “family state that”. Summaries tend to convert hedged statements into flat ones. Ask for uncertainty to be preserved, and check that it was.
Silent failure on scans. Referral packs are full of scanned and handwritten pages. Some tools will read them badly and some not at all, and most will not tell you which. If a pack includes handwriting, assume the tool did not read it.
How to trial it
- Take five referrals you have already processed, where you know what mattered.
- Summarise each with the tool.
- For each one, ask: would this summary have led me to the same decision? And specifically: did it include the thing that actually mattered?
- Count how often the answer is no. That is your error rate, and it is far more useful than minutes saved.
If the answer is no more than once in five, use it only for triage — deciding what to read first — and not for anything else.
What it costs
Assistant business plans in our directory start around £15 to £18 per user per month. Only a handful of people in most services handle referrals, so this is usually one of the cheaper uses to trial. See paying for AI.