AI scribes are arriving in clinics. Five questions to ask before you use one

The pitch is simple and, for anyone who has finished a clinic at eight and written notes until ten, very appealing. You put a phone on the desk, you talk to your patient, and by the time they have left the room a tidy draft note is waiting. These “ambient” scribes record the consultation, transcribe it and use a generative AI model to turn the conversation into structured documentation.

That last step is where the questions begin. A scribe is not a dictaphone. It hears a patient describing their symptoms, their family, their fears, sometimes things they have never said to anyone else, and it sends that audio somewhere to be processed. Before you switch one on, it is worth being as careful about the tool as you would be about a new drug: what it does, what it does to people other than you, and how it fails.

Here are five questions to ask, in roughly the order they matter.

1. Where does the recording go, and for how long?

Start with the plain facts of the data. Ask the vendor, in writing:

1. Is the audio stored at all, or deleted once it has been transcribed?
2. Where are the audio, the transcript and the draft note kept, and in which country?
3. Who at the company, or at any company it works with, can access them?
4. How long is each kept, and can you delete a patient’s data on request?
5. Is your patients’ data used to train or improve the vendor’s models?

These are not technicalities. India’s Digital Personal Data Protection Act, 2023 and its Rules, notified in November 2025, require organisations that process personal data to keep reasonable security safeguards, to tell affected people about a breach without delay, and to respond to requests to access, correct or erase their data.¹ If you run your own practice, you may be the organisation the law has in mind, not the vendor.

₹250 crore

The Rules are being phased in, with an eighteen-month period for organisations to comply.¹ That gives time to choose tools carefully. It is not a reason to postpone the question.

2. Did the patient agree, and could they say no?

Under the Act, personal data may be processed for a lawful purpose once a person has given consent, and that consent can be withdrawn at any time.² The Rules expect consent notices that are clear and set out the specific purposes for which data is collected.¹

In a consultation, that translates into something simple and humane: tell the patient the conversation is being recorded and why, in a language they understand, and make it easy for them to say no without feeling they have made your day harder. Some patients will not want to discuss certain things with a microphone on the desk. A good process lets you switch the scribe off for part of a consultation, or not at all, without fuss.

A scribe is not a dictaphone. It hears a patient describing their symptoms, their family, their fears.

Indian guidance on AI in healthcare puts informed consent and accountability among its core principles.⁵ Practically, that means writing down how consent is sought in your clinic, so that the answer does not depend on how busy the day is.

3. Who checks the note before it enters the record?

You do. Every time.

The World Health Organization’s guidance on AI in health sets out six principles for its use, among them protecting human autonomy and ensuring responsibility and accountability.³ Its later guidance on the large generative models behind tools like these makes the same point for this newer kind of system.⁴ In ordinary language: the software drafts, the doctor decides, and the doctor answers for what the record says.

That has a cost the sales material rarely mentions. Reading a draft carefully takes time, and a note that is fluent and confident is easy to approve without really reading it. The useful habit is to read the draft against your own memory of the consultation, before the next patient comes in, while you can still remember what was actually said.

4. How does it fail, and how would you notice?

Every tool fails somehow. The point is to know how. Generative models can leave things out, attach a statement to the wrong speaker, or produce a plausible detail that nobody said. Ask the vendor:

• What evaluation has been done, on what kinds of consultations, and can you see it?
• How does the tool handle the way your patients actually speak, including switching between languages within a sentence?
• What happens with a third voice in the room, such as a relative answering for the patient?
• When it is unsure, does it say so, or does it guess?

What to ask, and what a good answer sounds like

Question	A reassuring answer	A reason to pause
Is audio kept?	Deleted after transcription, with a stated time limit	“We keep it to improve the product”
Where is data stored?	A named provider and country, in writing	Vague, or “in the cloud”
Is our data used for training?	No, or only with explicit opt-in	Yes, by default
How was it evaluated?	Published or shareable results on comparable consultations	Testimonials instead of results
Can we delete a patient’s data?	Yes, with a documented process	Not without ending the contract

Then run your own small check. For the first few weeks, compare a sample of drafts with what you remember and what you would have written. If the errors are of a kind you would not reliably catch while tired, that is your answer.

5. What happens when you stop using it?

Contracts end, companies are bought and prices rise. Before you start, find out how you would leave: whether notes already in the record stay readable, whether audio and transcripts held by the vendor are deleted and how you would know, and whether you can export what you need in an ordinary format.

None of this is an argument against scribes. Time is one of the currencies a doctor’s life is short of, and a tool that returns an hour a day is worth taking seriously. It is an argument for choosing one the way you would want your patients to choose a doctor: by asking how it works, how it fails, and who is responsible when it does.