Welcome to the first issue of The Med-AI Brief.
Once a week, about five minutes, three things: what's actually real, one tool worth trying, and one practical way to save time. Evidence attached where it exists, and flagged where it doesn't.
What's real
Your colleagues are already using AI. Nobody told them how.
A survey of 1,000 NHS healthcare professionals found 90% use AI in their clinical work — 96% among hospital doctors, pharmacists and mental health professionals. The number that matters more: 65% got there ahead of any formal guidance from their workplace.
Caveat first, because it's a real one: that survey was commissioned by Heidi, a company that sells AI scribes. Treat the direction as credible and the decimal places as marketing.
Better evidence, less comfortable reading. Blease and colleagues surveyed 1,003 UK GPs for BMJ Health & Care Informatics (2026;33:e101847) — the largest UK study of ambient scribes so far. Among the 141 already using one:
80% spent less time on documentation, and 30% called the reduction large.
70% reported lower cognitive load during the consultation itself.
44% found errors in 10–30% of AI-generated documents.
14% had seen errors with significant to critical implications.
The errors cluster exactly where you'd fear: multiparty consultations (38%), complex histories (35%), and non-English consultations (31%). And 37% did not routinely ask patients for permission before recording.
What that means for you
The time saving is real — and the drop in cognitive load may matter more than the minutes.
So is the error rate, and it's your name on the record. "The AI wrote it" has never been a defence.
Read the output hardest when the consultation was hardest. Interpreter present, complex history, three relatives talking.
Ask the patient. One sentence. Skipping it is the fastest route from time-saver to complaint.
Adoption was 14% when the fieldwork ran in August 2025. Heidi has since signed a deal to roll out across NHS Midlands — the first region to procure ambient voice technology at scale — so that number is already history.
One tool worth trying
NotebookLM — for the reading pile, not the patient
NotebookLM answers only from sources you upload — guidelines, papers, a trust protocol — and cites back to the exact passage. That last part is the real difference from a general chatbot: you can check its answer in about five seconds.
Where it falls over: it only knows what you gave it, so it won't find literature for you, and it can still flatten a nuance. It's a consumer Google product — no patient-identifiable data, ever.
Try this: upload the last long guideline you meant to read and ask what changed from the previous version, and what that means day to day.
A quick win
Two minutes, in whatever assistant you already have open.
The anonymise-first letter prompt
Strip every identifier from your notes first — no name, DOB, NHS number, address. Use "a 62-year-old man". Then paste:
"Draft a clinical letter from a [speciality] clinic to a GP using my rough notes below. Use these headings: Reason for referral, Summary, Examination and investigations, Impression, Plan, Actions for the GP. Plain professional English. Do not add any clinical detail I have not given you — if something is missing, list it under 'Missing information' rather than inventing it. Notes: [paste]"
That last instruction earns its place. Without it the model fills gaps with plausible fiction; with it, you get a checklist of what you forgot to write down.
Then read every clinical detail before it goes anywhere near the record.
Regulator watch
UK. The MHRA published its AI Airlock Phase 2 report on 9 June 2026, covering seven technologies chosen from 51 applicants. Two findings matter clinically: generative systems can quietly "operate beyond their intended purpose", and statistical significance in validation doesn't mean clinical relevance. Phase 3 runs to 2029, feeding the AI framework the MHRA says it will publish this year.
Gulf. In Saudi, the baseline documents are the SFDA's MDS-G010 and MDS-G53 — one of the few countries with binding AI-specific device requirements rather than principles. Abu Dhabi and Dubai were among the earliest anywhere to regulate clinical AI. Practical translation: your facility's data policy, not the vendor's marketing, decides what goes into a tool.
That's issue one
Ambient scribes give real time back, and they get things wrong most often in exactly the consultations you'd least want them to. Use them. Read them properly. Ask the patient.
One favour: hit reply and tell me your speciality and where you work. I read every one, and it shapes what I write.
See you next week,
Saeed
Nothing here is clinical or legal advice. Check any tool against your own information-governance rules before it touches patient data.