MedFlow: AI intake that gave clinicians their mornings back
A working pipeline in the clinic — not a demo. 70% less admin before morning lists, with humans still accountable.
The situation
MedFlow’s operations team in the UK was drowning in inbound forms, attachments and incomplete histories. Vendors arrived with chatbot demos. Nobody owned evaluation, fallbacks or what happened when the model was wrong at 07:50.
What we did
We started with a risk register: which fields could be suggested, which required a person, where data lived, and how staff overrode the system. Then we shipped a supervised intake pipeline — classification, extraction, a review queue — into the workflow they already had, rather than a new island.
Training was part of delivery. If a nurse cannot see why a suggestion appeared, the feature is a liability. Logging made that visible.
Outcome
About 70% less admin before morning clinics on the measured workflow. The team kept the queue. We stayed for hypercare through the first live exceptions, then handed over a runbook.
See AI software