Questions to ask an AI scribe or assistant vendor, before you sign

The short answer

Picture a sales demo where the AI scribe writes a tidy note in seconds and the contract is ready to sign. The questions that matter come after the demo: where the recordings go, who can see the logs, and how you leave. Get the answers in writing before you sign.

Before you sign with an AI scribe or assistant, get written answers to six groups of questions: where patient data goes and whether it’s used for training; whether the vendor signs a business associate agreement (US) or a processor contract (India); how fast it tells you about a breach; how the draft is reviewed before it enters the record; what is logged and who can see the logs; and how you get your data out and deleted when you leave.

This is about systems and contracts, not clinical judgement. It doesn’t replace your own counsel’s review.

Why ask at all

Because the notes are not always right, and the contract decides who carries the consequences. A pilot published in 2026 that reviewed 356 notes drafted by an ambient AI scribe found omissions in 18%, hallucinations in 11.5%, and rated 5.3% as containing an error posing serious or imminent risk.¹ Our post on hallucination rates in clinical notes covers more studies. A vendor can’t make that risk zero. What you can control is review, records and the contract.

Data: where it goes, and what it’s used for

1. Where is audio and text processed and stored? Which country, which cloud provider, which region.
2. How long is the audio kept? Ask for the default and whether you can set it shorter.
3. Is any of our data used to train or improve models, yours or a third party’s? Get “no” in the contract if that’s what you want, not only in an FAQ.
4. Which sub-processors touch the data? Speech-to-text, the language model provider, hosting. Ask for the list and how you’ll be told of changes.
5. Who at the vendor can see patient data, and is that access logged?

The contract

In the US. A vendor that creates, receives, maintains or transmits protected health information on your behalf is generally a business associate, and that includes its subcontractors.² A business associate contract must, among other things, limit the vendor’s use and disclosure of PHI, require safeguards and the Security Rule’s protections for electronic PHI, require it to report breaches, bind its subcontractors to the same terms, and return or destroy PHI at termination if feasible.³

6. Will you sign our BAA, or show us yours? Read it for the items above.
7. How fast will you tell us about a breach? The rule’s outer limit for a business associate is 60 calendar days after discovery⁴, which is a long time for a clinic. Ask for days, in writing.

In India. Under the DPDP Act, the clinic or hospital stays responsible for processing a vendor does on its behalf, and may use a processor only under a valid contract.⁵

8. Will you keep processing logs for at least a year, as the DPDP Rules ask of processing done by a processor?⁶
9. Will you tell us about a breach in hours? The clinic owes the Data Protection Board a detailed report within 72 hours of becoming aware.⁶
10. Will you erase our data when we ask, and confirm it in writing?⁵

Review and errors

11. Does the draft ever reach the record without a clinician’s sign-off? The answer you want is no, by design, not by setting.
12. Can the reviewer see what the tool added versus what was said? Highlighting or a transcript side by side makes review faster and less error-prone.
13. Can the tool write anything other than the note, such as orders, codes, problem-list entries or messages to patients? If yes, which, and who approves each?
14. How do we report a wrong note, and what do you do with the report?
15. What changed in the last model update, and how are we told before the next one?

Question 13 matters most. A tool that drafts text is one level of risk; a tool that writes structured entries into the record is another. Our post on AI agents writing to patient records explains the difference.

Logs and records

16. Is every action the tool takes logged, with who, what, when and which patient?
17. Can we see those logs ourselves, without opening a support ticket?
18. Can the vendor edit or delete those logs? If yes, they can’t settle a dispute between you and the vendor.
19. Does the log keep the draft as generated and the note as signed, so you can show what the tool produced and what the clinician changed?

Exit

20. How do we get our data out, in what format, and how is the rest deleted? Ask for this in the contract, with a deadline.

A one-page scorecard

Copy this into a sheet and fill one column per vendor.

For the Indian contract terms in more depth, read AI vendors as data processors under DPDP. On why a quick approval click isn’t real review, see automation bias and approvals. For what Deedglass does around AI agent writes in hospitals, see /healthcare.

To try the checks on synthetic data, open the playground in your browser. To see what an AI agent on your machine can change, run npx deedglass scan (what the grade means) or paste its MCP config into /check; neither uploads your data. Our AI agent permissions report shows what public agent setups commonly leave open.

Written with AI assistance and checked against our claims register.