How an AI Scribe Gets a Note Wrong, and the Checks That Catch It
By Tanmay Mathur · Published 2026-10-07
The dangerous AI note is not the badly written one. It is the fluent one with a wrong dose in it.
Fluent is not the same as right
Language models write well. That is the problem: a note with a wrong dose reads exactly as confidently as a correct one. We test Clinic Scribe on hundreds of invented consultations in five languages, and every error below is one we actually found. Each now has a check that stops the doctor before signing.
The errors, and the checks
- A dose ten times too high. Speech recognition hears "Telma 40" as "Telma 400", or "Urimax zero point four" as 4. Both agree with the transcript, so only knowledge of the drug can catch them. A dose outside the usual range blocks signing.
- A changed dose kept at the old strength. "Serlift 50, increase to 75" - and the medicine list still says 50. Two passes over the conversation are compared; a disagreement blocks signing.
- A self-correction missed. "Betnovate - sorry, Mometasone." The corrected drug must win.
- A relative's medicine in the patient's list. A wife mentions her own Thyronorm. It is kept out and flagged by name.
- A "no" that disappears. "Fever?" "Nahi." The denial must reach the note, in any of the five languages.
- Spoons turned into millilitres. "Two spoons" becomes "2 ml", or "2 gm". Household measures stay as said.
- Half a number. "HbA1c seven point..." and the rest not heard, written as 7. The unfinished decimal is flagged.
- A follow-up that contradicts itself. "Six months, on 18 November" - 18 November is six weeks away. One of them was misheard or misspoken.
Why checks, not a smarter prompt
We tried telling the note model more about each specialty. It made the notes worse: given more medical knowledge, the model drifted toward what is typical instead of what was said. So the knowledge goes into checks - dose ranges for Indian brands, the negatives each specialty asks about - which cannot invent anything and can only stop a signature.
The doctor still reads every note
No check catches everything. The note shows the words behind every line, flags must be ticked one by one, signed notes cannot be overwritten, and the doctor reads every note. That is the design, not a disclaimer.
Frequently asked questions
Can an AI scribe make up something the patient never said?
It can try. Every line in a Clinic Scribe note sits beside the words it came from, and a claim with no supporting words is flagged before signing.
Does the scribe correct mistakes on its own?
No. It flags; the doctor decides. Nothing in the note is changed silently.
Related
Keep reading
- Can an AI Scribe Understand a Hindi or Telugu Consultation?
- When a Patient Says No to Being Recorded
- Patients Book and Then Do Not Turn Up. How to Fix No-Shows
See the checks in Clinic Scribe
If this sounds familiar
- We keep missing calls - what is that costing?
- Customers ring after we close
- The line is engaged all morning
- Should I hire a receptionist?
- People book and then do not turn up
Home · Pricing · Use cases · Industries · Comparisons · Products · Blog · Book a demo · All solutions · Hear a demo call
For doctors: Clinic Scribe, an on-premise AI medical scribe
Legal: Privacy Policy · Terms of Service · Cookies · Image credits · Security and trust · Acceptable use
{/* __FV_PRIVACY_V1__ */} {/* __FV_FOOTLINKS_V3__ /solutions/ had ZERO inbound links and it is the page that lists every comparePages entry, including all 25 service x city pages. */}AI receptionist for dental clinics · salons · coaching institutes · real estate agents · hospitals · diagnostic labs · AC repair · plumbers · electricians · appliance repair
Common questions: what it costs in India · does it speak Hindi · does it speak Telugu · does it speak Tamil · does it speak Kannada · does it speak Malayalam · does it speak Bengali · does it work with call forwarding · does it send a WhatsApp confirmation · can it transfer to a person · what if it does not understand
Compare: vs voicemail · vs a chatbot · vs forwarding to your mobile · Vapi alternative · Retell AI alternative · Bolna alternative · Exotel alternative · Knowlarity alternative · Ozonetel alternative · Gnani alternative · Yellow.ai alternative · SquadStack alternative · Haptik alternative · virtual receptionist India
{/* __FV_FOOTLINKS_V2__ these and the four new language pages were in Navbar.tsx only, which is React and never reaches a crawler - the same miss as __FV_CITY_CRAWL_V1__ */}AI receptionist by city: across India · Hyderabad · Bengaluru · Chennai · Delhi NCR · Mumbai · Pune · Kolkata · Ahmedabad · Jaipur · Surat · Lucknow · Indore · Bhopal · Nagpur · Chandigarh · Visakhapatnam · Vijayawada · Coimbatore · Madurai · Kochi · Mysuru