Client
Mersey Health Group
Sector
Healthcare
Headline
4 days · Cut from referral turnaround
Time to ship
Eight weeks
The problem
Mersey Health Group operates three primary care sites across the Liverpool City Region and was running a referral and clinical correspondence pipeline that had grown organically over a decade. Inbound referrals, lab results and consultant letters arrived in mixed formats — PDF, scanned paper, faxed images, attached to email — and were triaged manually by an administrative team. Average turnaround from receipt to patient record was seven working days. The clinical leadership team wanted that under three, but the obvious solutions — hiring more administrators, contracting a triage service — failed the cost test.
What we built
Two coupled systems sharing infrastructure: a document intelligence pipeline that reads inbound clinical correspondence and a retrieval-grounded internal assistant that helps the admin and clinical teams answer routine questions against the group's guideline library.
The document side: the pipeline reads the inbound document, identifies the patient (with a human reviewer for any ambiguity), classifies the document type, extracts the structured fields the clinical system needs, and routes the document and a structured summary to the responsible clinician for sign-off. Confidence threshold tuned conservatively; anything below it gets a human reviewer.
The assistant side: a retrieval-grounded chat assistant scoped tightly to Mersey Health's own guideline library, pathway documents and policies. Every answer cites the underlying document; anything outside the scoped corpus gets a refusal. Used by admin and clinical staff for the "what is our policy on X" questions that were eating senior time.
What changed
Measured against the pre-deployment baseline:
- Referral turnaround down from 7 working days to under 3 — a four-day reduction, the headline number we committed to.
- Document triage time down by 68% on the admin side; the team redirected the saved capacity into proactive patient follow-up.
- Clinician question-answering on policy down from 8–12 minutes per query to under 2 via the assistant, with citations on every answer.
- Zero clinical incidents attributable to the system in the first six months; every model-touched record passed a clinician review before being entered.
What we did not do
We did not build any system that touched clinical decision-making about individual patients. The pipeline structures inbound information and surfaces what needs attention; the clinician makes every decision. The line is important and easy to cross by accident in healthcare engagements; we drew it sharply at the start of the project and held it through delivery.
We also did not migrate Mersey Health's clinical system or core record-keeping. The new pipeline augments those systems via the existing integration endpoints. The clinical record stays in the audited, regulated system it has always lived in.
Compliance and audit
Models are hosted in a UK region; patient data does not leave UK infrastructure. Every model call, every retrieved guideline, every clinician sign-off is logged with a versioned timestamp. If a future clinical decision ever needed to be defended on how the system contributed, the audit trail exists end to end.
What is next
Mersey Health has a named internal owner for the system; we are on retainer for the next two quarters to tune as new document types are added. They are considering extending the same pipeline to inbound consultant letters from the wider NHS estate — a larger document corpus with more variation — which we will scope in the next quarter.
Related reading
For the broader pattern: AI for healthcare in Liverpool: where it is helping, where it is not.