Skip to content
HealthcareEnterpriseUK11 monthsProgrammeCS-2024-04 · disclosure level 4

Referral triage inside a boundary data cannot leave

Manual triage of inbound patient referrals, under a data residency rule that ruled out every hosted AI product.

Triage time
5.9 min52s-85%
Auto routed
0%68%+68pp
Misroutes
3.8%0.8%-79%

Six month observation window against a matched prior period. Source: internal triage audit, December 2024

Context

An enterprise healthcare provider receiving 3,000 referrals a week across 14 specialties, triaged by a clinical admin team reading free text letters.

The constraint

The thing that ruled out the obvious answer.

Patient data could not leave the organisation's network boundary under any circumstances. Every hosted AI product the trust had trialled failed procurement on that single clause, regardless of certification.

Diagnosis

Triage was not one decision, it was two. Classifying the specialty was mechanical and consistent. Judging urgency was clinical and contested, with genuine disagreement between reviewers on the same letter. Automating both together would have failed on the second and discredited the first.

Decision

What we rejected, and what we chose instead

Any firm can describe what it built. The useful record is the option that was on the table and did not survive.

Rejected

A single cloud model handling classification and urgency together

It failed procurement on residency, and it bundled a mechanical decision with a clinical one. Clinicians would have rejected the whole system because of the half that needed their judgement.

Chosen

A local classification model, with urgency left to clinicians

Runs entirely inside the boundary on existing hardware. It removes the mechanical half of the work and leaves the contested half where it belongs, which is why clinicians accepted it.

Build

What was actually implemented

PythonPyTorchOn premise GPUPostgresHL7 FHIR
  1. 01Fine tuned classification model running on the trust's own infrastructure, no external calls
  2. 02Confidence threshold below which a referral goes to a human reviewer unchanged
  3. 03Every automated decision logged with the input and the confidence score
  4. 04Shadow mode for eight weeks before any decision affected routing
  5. 05Clinical override always available, and overrides feed the next evaluation set
Result

Measured, not estimated

Median triage time fell from 5.9 minutes to 52 seconds on the 68% of referrals the model routed automatically. Misroutes fell by 79%, measured against the same audit process used before the programme.

Six month observation window against a matched prior period. Source: internal triage audit, December 2024

BEFOREAFTER-77%-55%-83%-34%
The routing model respected residency requirements from day one.
Programme Director, healthcare
Next

More from ai and automation

Professional services

The chatbot we talked them out of building

CS-2025-07

Pillar

AI and automation

Most AI projects fail because they add a chat box next to the problem instead of removing the step that costs the hours. We start from the workflow.