After a video consult, doctors retyped their own notes into every specialist referral letter: two to three extra minutes a case, with no template, so no two letters looked the same. Referrals are rare, about one patient in twenty, and doctors read every detail before sending.
- Problem
Retyping notes into every letter. - What I changed
An AI draft from the doctor's own notes. Every field editable. - Key result
30% of all referral letters are now AI drafted. - My role
Flow, AI interaction, motion, app handoff, design QA.
Two entry points explored, one shipped.
A prompt under the consult notes meets the doctor where they are writing. A button inside the existing referral drawer fits the flow doctors already use.
Button in the referral drawer. Chosen after review with the clinical lead.
Three decisions that kept the doctor in charge.
1. Regenerate button→Better notes, better draft
What changed
The button only works again when the notes change. Doctors improve the input instead of rolling the dice.
2. Greyed-out Save→Save that points to the gap
What changed
Save stays active and shows which field is missing, so nobody guesses why it will not save.
3. Four extra features→One drawer, one job
What changed
Cut the saved list, preview step, urgency field and bulk actions. Referrals happen once in twenty patients.
A referral drawer that drafts with the doctor.
Drafts only from the notes
The button stays disabled until diagnosis and notes are filled.
Then the AI drafts from what the doctor already wrote, nothing else.
A draft that says it is a draft
Every field stays editable, with a disclaimer that clears with the draft.
The doctor reviews and saves. Nothing is sent on its own.
Doctors keep the draft and rewrite the summary.
Measured from 1 Aug to 29 Oct 2026.
What I took from it.
- Designing AI for clinicians is mostly designing the edges. When it runs, what it reads, what happens when it is wrong, and how the person stays the author.
- Agree the success metric before launch. The brief named an 8% expired call rate on referral symptoms, and we have no after figure for it yet.
