AI in the Front Office

AI and Patient Data: What Vendors Do with Recordings

Retention, model training, deletion and breach terms to settle before an AI phone or chat vendor records a single patient call.

AssistCall research teamPublished Reviewed by Faruque Amin, VP of Sales and Marketing7 min read

On this page
  1. Why recordings need their own terms
  2. What one patient call can leave behind
  3. The business associate agreement comes first
  4. Eight terms to pin down in writing
  5. Model training: read the definitions
  6. Retention: set a number, not a promise
  7. Who inside the practice can listen
  8. Breach timelines your contract must support
  9. Questions for the vendor call
  10. Before the first real call
  11. Frequently asked questions
  12. References
AI in the Front Office: AI and Patient Data: What Vendors Do with RecordingsArticle coverAI IN THE FRONT OFFICEAI and Patient Data:What Vendors Do withRecordingsAva Medical IntakeInsightsAI in the Front Office: AI and Patient Data: What Vendors Do with RecordingsArticle coverAI IN THE FRONT OFFICEAI and Patient Data:What Vendors Do withRecordingsAva Medical Intake

Key takeaways

  • A single recorded call creates several copies: audio, transcript, summary, metadata, logs and backups. Each needs its own retention rule.
  • No business associate agreement, no protected health information. If a vendor will not sign one, design the workflow so health details never reach it.
  • Read the model training clause word by word. 'Improve our services' and 'de-identified' can both mean your calls train someone's model.
  • Write deletion, breach notice and subprocessor terms into the contract before go-live, not into a follow-up email.

Why recordings need their own terms

AI now plays some role in patient visits at 83% of medical groups, up from 71% a year earlier[1]. The paperwork has not kept pace. Only 20% of groups have AI governance or a formal AI-use policy, 22% are drafting one and 56% have none[2]. For front office tools, the gap usually shows up in one place: what happens to a call after it ends.

Medical groups with an AI-use policy, 2026
Medical groups with an AI-use policy, 2026Donut chart. Have AI governance or a formal policy: 20%; Developing one: 22%; No policy: 56%.Have AI governance or a formal policy: 20%Developing one: 22%No policy: 56%56% none
  • Have AI governance or a formal policy 20%
  • Developing one 22%
  • No policy 56%

Source: Medical Group Management Association (MGMA), 2026[2]

Show the data
ItemValue
Have AI governance or a formal policy20%
Developing one22%
No policy56%

Patients notice. In a 2026 poll, 77% of the public said they were concerned about the privacy of medical information given to AI tools[3]. Physicians share the worry: about 40% say they are equally excited and concerned about AI, and patient privacy is near the top of their concerns[4]. A practice that can say exactly where its recordings live, for how long and who can hear them has a straight answer when a patient asks.

What one patient call can leave behind

Vendors often describe 'the recording' as if it were one file. In practice a single call to an AI phone agent can produce all of the following:

  • Audio of both sides of the call. A voice is identifying on its own, and callers often say their name and date of birth.
  • A full transcript, sometimes in more than one language.
  • An AI-written summary or call note, which may be pushed into email, a ticketing tool or your practice management system.
  • Metadata: caller number, time, duration, routing path and outcome tags.
  • Structured fields the system pulled out, such as name, callback number or reason for call.
  • Debug and quality logs held by the vendor and its subprocessors.
  • Copies reviewed by the vendor's staff or contractors for quality checks.
  • Backups, which often outlive the original by months.

Each copy can sit in a different system with a different retention period. Your contract has to cover all of them, not just the audio file you can see in the dashboard.

The business associate agreement comes first

A covered practice may let a vendor create, receive or keep protected health information only after it gets satisfactory assurance, in a business associate agreement, that the vendor will safeguard it[5]. For an AI phone or chat tool that records patients, that agreement is the entry ticket, not a nice-to-have. Check that it covers the vendor's subprocessors, limits use to the services you contracted for, and requires return or destruction of data when the contract ends.

If a vendor will not sign one, the honest options are to walk away or to design the workflow so health details never reach the tool. Some front office tools are built that way on purpose. Ava, for example, answers office questions from wording the practice approves and passes anything about a specific patient to staff rather than taking health details. Whatever you choose, test it with fictional calls and confirm that callers who volunteer symptoms are cut off politely and routed, not transcribed in full.

Eight terms to pin down in writing

Contract terms for any AI tool that records or transcribes patient calls
TermQuestion to askAnswer to acceptAnswer to push back on
Audio retentionHow long is call audio kept, and can we set it?A fixed period you choose, including zero'As long as needed to provide the service'
Transcripts and summariesDo they follow the same retention as audio?Yes, or a shorter period you setKept indefinitely for analytics
Model trainingIs our data used to train or fine-tune any model?No, stated in the agreementOnly 'de-identified' data, with no definition
Human reviewWho listens to calls, where, and why?Named roles, for support you requestContractors for general quality checks
SubprocessorsWhich companies touch our data?A published list with advance notice of changes'Industry-standard partners'
Deletion at exitWhat is deleted when we leave, and when?All copies, including backups, by a set date, with written confirmationBackups age out 'eventually'
Breach noticeHow fast will you tell us about an incident?A specific number of days, well inside the legal deadline'Promptly'
ExportCan we export recordings and transcripts?Yes, in a standard formatOnly through a paid project

Model training: read the definitions

Most disputes come down to definitions. A clause saying the vendor will 'use customer data to improve the services' can cover training a shared model on your calls. A promise to train only on 'de-identified' data means little unless the agreement says de-identified under the HIPAA standard, who does it, and whether voice audio is ever included. Ask for three plain statements in the agreement itself, not the marketing page:

  1. Our audio, transcripts and summaries are not used to train or fine-tune any model, including the vendor's own.
  2. Any AI engine provider the vendor uses is contractually barred from retaining or training on our data.
  3. Opting out of training costs nothing and does not reduce features.

Retention: set a number, not a promise

HIPAA requires covered entities to keep required policies and compliance documentation for six years[6]. That rule is often misread as a reason to keep call audio for six years. It is not. Keep your vendor contract, your BAA and your retention policy for that period, and keep the recordings themselves only as long as they serve a purpose you can name.

Example retention schedule for a fictional four-provider practice: call audio 30 days, transcripts 90 days, AI summaries copied into the chart follow the medical record policy, metadata 13 months for staffing analysis, vendor debug logs 14 days, backups purged within 35 days of the source record.

Write your schedule into the order form and ask the vendor to confirm in writing that its systems enforce it automatically. Then test it: pick a fictional test call, note the date, and check that it is gone when it should be.

Who inside the practice can listen

Vendor terms are only half the picture. The other half is your own access settings. A dashboard that lets every staff login play every call turns a useful tool into a privacy problem. Set it up the way you would set up chart access:

  • Give playback rights to the practice manager and named supervisors only. Front desk staff usually need summaries, not audio.
  • Turn on the access log and check it monthly. You want to see who played which call and when.
  • Switch off downloads unless someone has a written reason, such as a complaint review.
  • Remove access on a staff member's last day, in the same checklist as their EHR and email accounts.
  • Decide who may share a recording outside the practice, for example with a patient or your attorney, and log every time it happens.

Recording consent rules differ by state, and some require every party on the call to agree. Keep the spoken notice short and place it at the start of the call, before the caller shares anything personal.

Breach timelines your contract must support

If a breach exposes patient information, the practice must notify affected people without unreasonable delay and no later than 60 calendar days after discovery[7]. Breaches affecting fewer than 500 people must be logged and reported to HHS within 60 days after the end of the calendar year[8]. You cannot meet either deadline if your vendor takes weeks to tell you. Penalties start at $145 per violation in the lowest tier and run to an annual cap above $2.19 million per identical provision[9], so a vague notice clause is an expensive place to save negotiating time.

Ask for a fixed number of days for the vendor to notify you, a named contact on both sides, and a duty to share the facts you need: what data, how many patients, which dates, and what the vendor has already done.

Questions for the vendor call

  1. Map the data

    Ask the vendor to draw every copy a single call creates and where each one lives. If they cannot, that is your answer.

  2. Ask for the documents

    Request the BAA, subprocessor list, retention settings screen and security summary before the demo ends.

  3. Test the edge cases

    Ask what happens when a caller reads out a card number, describes symptoms or asks to stop being recorded.

  4. Confirm the exit

    Ask how you get your data out and how deletion is proven. Get the answer in the contract.

  5. Record the decision

    Log who approved the tool, which terms were agreed and when you will review them, as part of your AI use policy.

Before the first real call

  • BAA signed, or a written design decision that no health details reach the tool.
  • Retention periods set in the vendor settings and screenshotted for your file.
  • Training opt-out confirmed in the agreement.
  • Recording notice wording approved for your state's consent rules.
  • Staff told how to answer 'Is this call recorded, and who hears it?'
  • A review date on the calendar, at least yearly, or sooner if the vendor changes subprocessors.

Done once, this work makes every later AI decision faster, whether you are adding overflow answering at peak times or comparing tools with a vendor evaluation checklist.

Frequently asked questions

Can an AI phone vendor record patient calls without a BAA?

Not if the calls include protected health information. A vendor that creates, receives or keeps that information on your behalf needs a business associate agreement first. Without one, the tool must be set up so callers' health details are not captured.

How long should we keep AI call recordings?

Only as long as you have a named purpose, such as quality review or dispute handling. The six-year HIPAA rule applies to policies and compliance documents, not to every recording. Many practices keep audio for weeks, not years.

Is de-identified data safe to let a vendor train on?

Only if the agreement defines de-identification to the HIPAA standard and excludes voice audio, which can identify a person by itself. If the definition is missing, treat the clause as permission to train on your calls.

What should we tell a patient who asks about recordings?

Say whether the call is recorded, why, how long it is kept and who can listen. Keep a short approved answer at the front desk so every staff member gives the same one.

References

  1. Medical Group Management Association (MGMA) (2026). As AI use expands, medical practice leaders shift attention to measurement (opens in a new tab) ↑

  2. Medical Group Management Association (MGMA) (2026). AI governance in medical group practices: Rules for the humans in the loop (opens in a new tab) ↑

  3. KFF (2026). KFF Tracking Poll on Health Information and Trust: Use of AI For Health Information and Advice (fielded Feb 24 - Mar 2, 2026) (opens in a new tab) ↑

  4. American Medical Association (2026). More than 80% of physicians use AI professionally: AMA survey (opens in a new tab) ↑

  5. Legal Information Institute, Cornell Law School (Electronic Code of Federal Regulations text) (current). 45 CFR § 164.502 - Uses and disclosures of protected health information: General rules (opens in a new tab) ↑

  6. Legal Information Institute, Cornell Law School (Electronic Code of Federal Regulations text) (current). 45 CFR § 164.530 - Administrative requirements (opens in a new tab) ↑

  7. Legal Information Institute, Cornell Law School (Electronic Code of Federal Regulations text) (current). 45 CFR § 164.404 - Notification to individuals (opens in a new tab) ↑

  8. Legal Information Institute, Cornell Law School (Electronic Code of Federal Regulations text) (current). 45 CFR § 164.408 - Notification to the Secretary (opens in a new tab) ↑

  9. Legal Information Institute, Cornell Law School (Electronic Code of Federal Regulations text) (current). 45 CFR § 102.3 - Penalty adjustment and table (opens in a new tab) ↑

Book a demo

See Ava handle your calls.

Book a short demo. We’ll walk through your calls and workflow, and agree what Ava should take care of.

Book a demo