Healthcare · Aeris Health
An AI triage assistant clinicians actually trust
A hospital group wanted AI in its intake process without putting a black box between patients and care.
- Less documentation time
- 54%
- Clinician override rate
- 6%
- Recommendations with citations
- 100%
The challenge
Intake nurses were spending 40% of their shift on documentation. Off-the-shelf AI tools were rejected in pilot because clinicians could not see why a recommendation was made.
What we did
- 01Built the evaluation harness first, with clinician-labelled cases as ground truth
- 02Designed a retrieval pipeline over the group's own protocols, not general web data
- 03Every recommendation cites the source protocol paragraph it came from
- 04Added an escalation path: low confidence routes to a human, always
- 05Ran a six-week shadow deployment before a single output reached a patient record
The outcome
Documentation time per intake fell by more than half, with a clinician override rate of 6% — and a full audit trail that satisfied the group's clinical governance board.