Up to eight hours per clinical report, 400 reports a month. Now under fifteen minutes.
The problem they handed me
Every diagnostic report had to be checked line by line against the group’s clinical rulebook before it could go to a family. Senior reviewers were spending up to eight hours on one report, the backlog grew with every new clinic, and nobody believed AI could be trusted with clinical data. The board wanted proof before budget. The DPO wanted to know where the data would go.
What happened next
-
Week 1
Use-case discovery across the group. This case chosen on value and feasibility. Requirements agreed with QA, clinical leads and the executive team.
-
Week 2
Proof of concept on 15 real reports. Each report was reviewed manually and by the system, and every issue flagged was judged valid or not by the lead QA manager and lead clinician. The system found more valid issues than the manual review. Buy-in won on results, not slides.
-
Week 3
Business case approved by the executive team: forecast savings of up to £300K a year across two businesses, with benefits tracking set up to report measured savings from 2027.
-
Week 3
The blocker solved. Clinical data could not go to AI models processed in the US. I found and validated a route that keeps every AI call inside the UK, hosted on Azure UK South with private networking and single sign-on. Designed with the DPO, signed off before go-live.
-
Weeks 4–9
Production build. Three independent AI reviews per report against a versioned, clinically approved rulebook, then a deterministic gate that fails closed before any comment reaches a reviewer. Rulings that let QA managers override the AI and bind every future report (180+ in pilot). A tamper-evident audit trail of every model call and decision. Penetration tested.
-
Go-live
Live. 10 reviewers trained. Rulebook and go-live criteria signed off by clinical and QA leads. The QA lead co-built it and now trains reviewers without me.
-
Now
Rollout to the second group business under way.
Results
- Review time
- Up to 8 hours → under 15 minutes
- Hours freed up
- An estimated 1,500 a month: 400 reports a month × about 3¾ hours saved on each (an average of around 4 hours down to under 15 minutes)
- Forecast savings
- Up to £300K a year across two businesses
- Pilot accuracy
- 73 reports, ~4% of AI comments withdrawn by reviewers
- Benchmark
- 41 of 46 seeded defects found; 31 of 33 clean controls left alone
- Run time and cost
- Median under 60 seconds, about £0.30 per report
For you, this means: if your most expensive people are doing work a system should do, and your compliance team is nervous about it, I have already taken that exact problem from a sceptical board to a live system with a signed-off audit trail. I can do it again on yours, and you will see the proof in two weeks.