Case study · Remove the bottlenecks

An hour a letter, down to five minutes.

BZ Clinic's specialist medical typist was retiring, and replacing her was going to be hard and expensive. We built a custom AI tool for clinical practice into the clinic's existing workflow instead: the clinician dictates as normal, and a fully formatted patient letter comes back in minutes. It's live today.

No sensitive patient data goes into anything we build. It formats, it doesn't decide.

5 min down from an hour
per letter
Live today, and in daily use at the clinic.
The problem

A retiring typist, and a strict privacy constraint

BZ Clinic is a neurodevelopmental practice in Surrey, run by Dr Bozena Zoric with a small admin team. Every clinic letter, diagnosis, medication, risk, next steps, followed a strict clinical template and was produced by a dedicated medical typist who was retiring in mid-2026. Replacing a specialist medical typist is hard, and it's expensive.

The letters cover child health records, so the constraint was non-negotiable: no patient names, dates of birth, addresses or GP details could ever reach an AI tool. Whatever we built had to make that true structurally, not as a rule someone had to remember to follow.

Clinic letters and paperwork on a desk
The workflow

Three layers, and identifying data never touches the middle one

Privacy is structural here, not procedural. The layer that does the AI drafting never sees a patient's name, date of birth or address, because it's never given them in the first place.

A clinician writing patient notes by hand
01

Dictation

Same routine, no patient names

Dr Zoric dictates straight into the tool, the way she always has. She uses the patient's first name only, and the tool removes even that before anything reaches the AI. The recording is transcribed on our own UK server and deleted as soon as the text exists.

AI drafting

Transcript in, formatted letter out

The tool turns the transcript into a finished letter in the clinic's own template, structure, headings and tone, with the letterhead, signature and approved standard wording already in place. It auto-detects the letter type (new patient letters to parents, or follow-up letters after a review), handles gender-specific content and medication side-effect text with BNF links, and ends every letter with a checklist of anything for the admin team to verify.

Details + QA

Details added, then approved as before

The admin team adds the patient's identifying details from Cliniko, the practice system. Dr Zoric reviews and approves, and the letter goes out exactly as it always did.

The build

Trained, not just built

Across five iterative rounds, we tested the tool's output against real, anonymised patient letters, including Dr Zoric's own red-text corrections, extracted programmatically so every fix fed straight back into the next draft. The retiring typist's own letters set the structural authority throughout. Built and handed over in under a month, ready before their last day.

Round 1

Baseline draft, tested against the retiring typist's own letters as the structural template.

Round 2

First correction pass, using Dr Zoric's real red-text edits on a fresh batch of anonymised letters.

Round 3

Refined again against further real, anonymised letters.

Round 4

Consistency check across the new patient and follow-up letter types.

Round 5

Final validation round, then training for the admin team to run it day to day.

Then we made it a proper product. The first version ran inside off-the-shelf tools. In August we moved it onto our own secure, UK-hosted platform: dictation built in, letters produced with the clinic's own letterhead and signature, and every letter deleted from our side after 24 hours. The clinic now pays a monthly subscription, and it costs less than the specialist typist did.

What changed

Live today, in daily use

~92% less time per letter
<1 month built and handed over, ready before the typist left
“Excellent.”
Dr Bozena Zoric · on the first test letter
Said plainly

No identifying patient data ever touches the AI. That's structural, not procedural: the layer that drafts the letter is never given a name, date of birth, address or GP detail in the first place, so there's nothing for it to leak. GDPR-safe by design, not by policy.

Where this fits

This is what "removing the bottlenecks" looks like

Remove the bottlenecks

See the full AI infrastructure service this case study is the proof of. See how it works →

Where is your team losing hours?

One call, and we'll tell you honestly whether something like this would help your clinic.

Book a discovery call →
Book a discovery call →