Specialist technology and software for state and local government info@hqt.orgCommunitiesGlossarySearch

AI & Automation

AI scribes and ambient documentation in agency health clinics

How public health, behavioral health and correctional clinics should evaluate AI scribes: review, consent, recordings, and contract terms.

Ambient documentation tools listen to a clinical encounter, transcribe it, and produce a draft note in roughly the format a clinician would otherwise type. Vendors now pitch these products to public health departments, county behavioral health agencies, school based health centers, and correctional health programs, usually on the promise of less after hours charting. The clinical case is often reasonable. The questions that stall the purchase are rarely clinical. They are about what happens to the recording, who can later obtain the draft, and whose signature stands behind the note.

Agency clinics carry obligations a private practice does not. They may hold substance use disorder records covered by federal confidentiality rules, operate under court oversight or a consent decree, answer to a state public records act, and serve patients who speak languages other than English or who are in custody and cannot freely decline. None of that rules out ambient documentation. It does determine how the workflow and the contract have to be written.

What the technology does well, and where it stops

  • Reliable strength. Turning conversational speech into structured narrative, especially history and subjective sections, and carrying forward context that was spoken aloud.
  • Blind spot. Anything not spoken. Chart review, prior results, and reasoning the clinician never says out loud do not appear in the draft.
  • Error prone without review. Numbers said quickly, such as doses, vitals and dates; sound alike medication names; negation, where "denies" and "reports" invert meaning; and attribution when several people are speaking.
  • Invention. Generative drafting can produce fluent clinical language nobody said, including normal exam findings that were never performed. Treat every draft as unverified until a clinician reads it line by line.

Settle these points before the pilot, not after the first complaint.

  • Decide when the draft enters the EHR, and whether an unsigned draft held in the vendor's system is itself an agency record.
  • Require clinician review and signature. Do not permit automatic filing or auto signature, and do not let a queue of unreviewed drafts accumulate.
  • Write attestation language with counsel and your health information management lead. Many agencies note that documentation was AI assisted and that the signing clinician reviewed and corrected it.
  • Keep an audit trail of what changed between draft and signed note. It is valuable for quality review, and it is also discoverable.
  • Decide whether transcripts and audio belong to the designated record set, which affects patient access and amendment requests under HIPAA.

A business associate agreement is the floor, not the analysis. Several other regimes commonly apply in agency settings, and the details change, so confirm current requirements with your counsel and privacy officer.

  • Part 2 records. Federal rules for substance use disorder treatment records carry redisclosure restrictions and limits on use in legal proceedings that go beyond HIPAA. Recent federal rulemaking has moved Part 2 closer to HIPAA in some respects, but not identical. A vendor that treats everything as ordinary PHI may not satisfy Part 2.
  • Behavioral health. HIPAA gives psychotherapy notes distinct handling, and many states add their own mental health confidentiality statutes. Ask what a recorded therapy session produces and where it lives.
  • Recording consent. Some states require consent from all parties to record a conversation. Build a spoken script and posted notice, offer a refusal that is honored without delaying care, and document refusals.
  • Custodial settings. Consent in jails and prisons deserves extra scrutiny. Consider who else is present, including officers, and what side conversation the microphone captures.
  • Employees. Clinicians are being recorded too. Where staff are represented, check notice and bargaining obligations before deployment.

Accuracy, terminology, and non-English encounters

Ask for evidence by language and by setting, not a single global accuracy figure. Performance differs sharply between a routine primary care visit and a crisis encounter with overlapping speech.

Test with your own vocabulary: street names for substances, local slang, program acronyms, and the medication list your formulary actually uses. For encounters in other languages, test the real configuration, which usually means a three way conversation with an interpreter, plus code switching mid sentence. An ambient tool is a documentation aid and is not a substitute for a qualified interpreter; language access duties that attach to recipients of federal financial assistance generally continue to apply. Ask what the system does when confidence is low, and whether it flags uncertainty or silently guesses.

Recordings, retention, and public records

  • Require a data flow description: where audio goes, which subprocessors touch it, whether any processing happens outside the United States, and whether audio ever leaves the vendor's environment.
  • Prohibit use of recordings, transcripts, and notes for model training, tuning, or product improvement without separate written authorization, and extend the prohibition to de identified and derived data.
  • Set retention and deletion terms in the contract, with deletion certified on request and on termination. Confirm the schedule matches your records retention schedule and involve your records officer.
  • Make sure litigation holds and subpoenas can reach vendor held data. Automatic deletion and a preservation obligation can collide.
  • Expect public records questions. Medical records are commonly exempt from disclosure under state law, but exemptions are state specific and rarely cover configuration files, audit logs, contracts, and vendor correspondence.

What to ask vendors and put in the contract

  1. Show the full data path for audio, transcript, and draft, including retention at each stage.
  2. Will you sign a BAA, and can you meet Part 2 obligations where those records are involved?
  3. Provide accuracy evidence by language, by specialty, and for encounters with interpreters.
  4. Describe error modes you have observed, including fabricated content, and how the interface surfaces uncertainty.
  5. Contract for: no training on agency data, deletion with certification, breach notice timelines, audit rights, export of agency data on exit, and the right to disable specific features.
  6. Treat security attestations as supporting evidence, not as a substitute for your own review.

How to start

Pilot in one clinic with a defined baseline. Track how much clinicians edit drafts, not just how much time they save. Bring counsel, the privacy officer, the records officer, health information management, and frontline clinicians into the design before go live, and write the consent script before the first recording. Keep a documented path to turn the tool off in a specific program if accuracy or consent problems appear.

HQT works with state and local agencies on technology evaluation, integration, and governance for systems like these.

Frequently asked questions

Does an AI scribe vendor need a business associate agreement?

If the vendor creates, receives, maintains, or transmits protected health information on behalf of a HIPAA covered agency clinic, a business associate agreement is generally required. A BAA alone does not cover every obligation, since substance use disorder records, state mental health confidentiality laws, and recording consent statutes may impose additional terms. Have your privacy officer and counsel confirm which regimes apply to each program before signing.

Can the draft note an AI scribe produces be requested under a public records act or in litigation?

Assume that drafts, transcripts, and audit logs can be sought in discovery, and that state public records analysis is fact specific. Medical records are commonly exempt from public disclosure under state law, but related material such as configuration settings, contracts, and vendor correspondence may not be. Work with counsel and your records officer to classify each artifact and set a retention schedule before deployment.

Who is responsible for errors in an AI generated clinical note?

The clinician who reviews and signs the note is responsible for its content, which is why automatic signing should be disabled. Build attestation language stating that the note was AI assisted and reviewed by the signing clinician, and keep an audit trail of edits between draft and signature. Treat unreviewed drafts as unverified material, not as documentation.

How should a clinic handle patients who do not want to be recorded?

Provide clear notice before the encounter, use a spoken consent script, and make refusal simple and free of any delay or reduction in care. Document the refusal and have a fallback documentation workflow ready. In custodial settings and in encounters involving minors or guardians, review with counsel whether consent can be considered voluntary and who must give it.

Can an AI scribe replace an interpreter for non-English encounters?

No. Ambient documentation tools are documentation aids, and language access obligations that apply to recipients of federal financial assistance generally require qualified interpretation. If you plan to use a scribe during interpreted visits, test the tool in that exact configuration, including three way conversation and code switching, and require the vendor to provide evidence of performance by language.

Get in touch

Have a project or a solicitation coming up?

Talk with an HQT engineer. We respond to agency, corporate, and partner inquiries promptly.