Key points
- Ask one question in writing. Will our patients' audio, notes or messages be used to train or improve any model, the supplier's or a third party's.
- The ICO says a provider that decides how to create and train its own model is likely to be a controller for that processing.2
- HIPAA's Safe Harbor method lists 18 identifier types to remove for de-identification, including voice prints and full face images.3
- The FTC has warned that quietly changing terms to allow AI training could be unfair or deceptive, so the commitment belongs in the contract.4
- wai never uses patient data to train models, and Profile data is not sold and not used to train external models.56
Summary
Get a contract commitment that patient data, including audio, notes and messages, is never used to train or improve any model. Where de-identified data is offered as the basis for training, ask which legal method is used and who documented it.
You tell every patient, in one plain sentence, that their consultation serves their care and nothing else. That sentence rests on a training commitment agreed before you sign.
See it in your clinicThirty minutes on your own patients and doctors.
Request a demoSee where your patients' data goes inside wai, from the consultation to the signed note, with the subprocessor list and the training commitment in front of you.
| Question to ask | Answer to look for | Follow up when you hear |
|---|---|---|
| Is our data used to train or improve any model | No, and it is written into the contract | Only de-identified data, or only to improve the service |
| Is consultation audio kept after the note is drafted | No, or only for a period the clinic sets | Kept for quality purposes with no stated limit |
| Which subprocessors see patient data | A named list with roles and notice of changes | A general description of trusted partners |
| Can the terms on data use change | Only with written notice and a right to leave with a full export | Terms may be updated from time to time |
| What happens to our data when we leave | Full export, then deletion with written confirmation | Deleted in line with internal policy |
| Which de-identification method is used, if any | Named method, named assessor, documentation shared | Industry-standard anonymisation |
The question to settle first
Every AI tool sees patient data to deliver its service. Ask whether any of that data then becomes training material for a model.
Data used only to deliver the service stays under the clinic's control and can be exported or deleted. Once data trains a shared model, the model keeps what it learned.
In the AMA's 2026 survey of US physicians, 86% rated data privacy assurances important for adopting AI.1 Concern about patient privacy reached 71% for AI tools not provided by the physician's own institution, against 42% for tools the institution provides.1
Training changes who is responsible for the data
Under UK GDPR the clinic is the controller of its patient records and the AI supplier is normally a processor acting on its instructions. The ICO says a provider that decides how to create and train the model behind its service is likely to be a controller for that element, while remaining a processor for the outputs it returns to each client.2
Where a provider keeps each client's models isolated, the client decides whether its own data improves its own model, and the isolation is complete and auditable, the client stays the sole controller.2
Training shared models on patient data needs its own lawful basis and a condition for processing health data, and patients must be told. A supplier that keeps clinic data out of training avoids that second use entirely.
What de-identified means in law
Under HIPAA, health information is de-identified by one of two routes. An expert applying accepted statistical methods determines and documents that the risk of re-identification is very small, or a fixed list of identifiers is removed and the organisation has no actual knowledge that what remains could identify the patient.3
The Safe Harbor list runs to 18 types, from names and dates to phone numbers, record numbers, biometric identifiers including voice prints, and full face photographs.3 A consultation recording carries a voice, and a free-text note often carries dates and family details.
Ask which method is used, who made the determination, whether audio is kept after the note is drafted, and whether the clinic can see the documentation, all in writing.
Contract terms that settle it
In February 2024 the US Federal Trade Commission warned that adopting more permissive data practices, such as using data for AI training, and telling people only through a surreptitious, retroactive change to terms or a privacy policy could be unfair or deceptive.4
Write four clauses. Clinic data, including audio, transcripts, notes and messages, is never used to train or improve any model. It is never sold or shared beyond delivering the service. Any change needs written notice and a right to leave with a full export. At the end of the contract all clinic data is deleted with written confirmation.
Ask for the subprocessor list, and check that any model provider's terms on training and retention match the promise. More questions sit in our list of questions for an AI scribe vendor.
What to tell patients
A clinic whose AI suppliers keep patient data out of training has a short, plain answer for the consent conversation and the privacy notice. Our guide to patient consent for AI scribes covers the wording.
The same standard follows the patient's own record. With wai Profiles, the patient's data is not sold and not used to train external models.6
wai's position
wai's pricing page states it in one line. Patient data is never used to train models.5 Every subprocessor is named with what it does and the data processing agreement in place, with thirty days notice before anything changes.7
wai hosts each clinic's records in its own country, under that country's rules. We set this up with your clinic. Notes and records can be exported, so the data stays the clinic's to keep.
Where to start
Send the training question in writing to each shortlisted supplier, with the four clauses attached, and ask for the subprocessor list. Record each answer beside that tool in the clinic's AI tool register.
Where an answer relies on de-identification, ask for the named method and the documentation before you proceed. Write the agreed clauses into the contract before go-live and file it with the procurement checklist.
Bring your current supplier contracts and see wai's training commitment, subprocessor list and export.
Frequently asked questions
May an AI supplier train on patient data?
It depends on the supplier's role, the lawful basis, any de-identification and the contract. The ICO notes that a provider training its own model is likely to be a controller for that processing, with its own obligations.2
Is de-identified data safe to use for training?
It can be, when done properly and documented. Under HIPAA, Safe Harbor requires removing 18 identifier types including voice prints, or an expert determines and documents that the risk of re-identification is very small.3
Does improving the service count as training?
It can. Ask for a definition of improve, and whether any model learns from your data. The answer to look for is that clinic data is used only to deliver the service to that clinic.
Can data terms change after we sign?
Put the commitment in the contract with a notice period and a right to leave. The FTC has warned that quietly changing terms to allow AI training may be unfair or deceptive.4
Does wai use our patient data to train models?
Patient data is never used to train models.5
What should we tell patients?
That the clinic's AI tools draft for a clinician or a member of staff to check, that their data is not used to train models, and where to read the privacy notice.
Sources
- 1American Medical Association, Center for Digital Health and AI, 20262026 Physician Survey on Augmented IntelligenceRead 14 October 2026Read source
- 2Information Commissioner's OfficeWhat are the accountability and governance implications of AI?Read 14 October 2026Read source
- 3Electronic Code of Federal Regulations45 CFR 164.514, other requirements relating to uses and disclosures of protected health informationRead 14 October 2026Read source
- 4US Federal Trade Commission, Office of Technology, 2024AI (and other) Companies, Quietly Changing Your Terms of Service Could Be Unfair or DeceptiveRead 14 October 2026Read source
- 5wai, 2026PricingRead 14 October 2026Read source
- 6wai, 2026wai Profiles, your personal health recordRead 14 October 2026Read source
- 7wai, 2026Security at waiRead 14 October 2026Read source
Your patients' data, working only for your clinic.
Fifteen minutes with a clinician who uses wai daily, with the training commitment and subprocessor list on the table. We reply within one working day.
Partner with us
See it on your own patients, then build it with us
Thirty minutes on your own patients, doctors and rules, in your market. We reply within one working day, and the clinics that partner with us early shape what we build next.
See what wai does between the visits