The Compliance Checklist for AI Conversations Across Regulated Industries

If your AI agent answers for a medical practice, an insurance agency, or a lender, the compliance surface is not a footnote to the project — it is the project. The good news: the obligations rhyme across industries. Consent, disclosure, retention, auditability, and a hard line around licensed activity are the same five questions every regulator asks, wearing different acronyms. Here is one checklist that covers all three.

This is deliberately not legal advice, and it is not a substitute for your own counsel or your regulator’s current guidance. It is an operator’s pre-flight list — the controls you should be able to point to before an AI agent sends its first message on your behalf. Work it top to bottom. If you can’t check a box, that’s your next sprint.

1. Consent: get it, log it, honor it

Every regulated conversation starts with permission to have it. For outbound SMS and calls in the U.S., that means express written consent under the TCPA and 10DLC registration for your messaging traffic — the baseline our compliance playbook covers in depth. On top of that, healthcare adds HIPAA’s requirement that you have a relationship and a permissible purpose before you discuss anything protected, and lending adds its own rules about when and how you can contact a consumer.

The failure mode is rarely “we never got consent.” It’s “we got it and can’t prove it,” or “someone opted out three weeks ago and the agent messaged them anyway.” An AI agent sends at machine speed, so a stale suppression list becomes a machine-speed violation.

• Capture consent with a timestamp, the exact language shown, and the channel it was granted on — and store it where an auditor can retrieve it in one query.
• Honor STOP, opt-outs, and revocation in real time, across every channel, so a text opt-out also stops the voice follow-up.
• Re-check consent state at send time, not at campaign-build time. The gap between the two is where violations live.

2. Disclosure: say it’s AI

Telling people they’re talking to an AI is moving from courtesy to law. As of 2026 several U.S. states require clear disclosure of automated or AI agents in certain interactions, and the direction of travel is one way. Beyond the letter of any single statute, disclosure is the cheapest trust you will ever buy — and the data says customers want it.

The wariness is real: Zendesk’s CX research finds a majority of consumers want companies to be more transparent and careful about how they use AI in support (vendor-published). Read that as a mandate, not a warning. Disclose early, make the human escape hatch obvious, and the same customers stop objecting.

• Identify the agent as AI at the start of the conversation, in plain language, not buried in a footer.
• Make “talk to a person” always available — no maze, no three-failed-loops gate.
• Keep the disclosure language versioned, so you can prove what was shown on any given date.

3. Retention: keep what you must, purge what you must

Regulated data has two clocks running at once: a minimum you’re required to keep and a maximum past which holding it becomes a liability. HIPAA-adjacent records, insurance communications, and lending disclosures each carry retention floors measured in years; privacy regimes and data-minimization principles push the other way. Your AI transcripts are records — treat them like the rest of the file, not like chat logs.

• Set retention by record type and jurisdiction, applied automatically — not by someone remembering to run a cleanup.
• Encrypt transcripts at rest and in transit, and scope who can read them to the people who need to.
• Support legal hold and defensible deletion. “We deleted it” should be provable, and so should “we preserved it.”

4. Auditability: assume you will be asked to prove it

The question an auditor, a regulator, or a plaintiff’s attorney actually asks is narrow: show me exactly what this system said to this person on this date, who could see it, and what changed.If your answer involves stitching together screenshots and Slack messages, you’ve already lost the afternoon.

An AI conversation you can’t reconstruct on demand is a liability wearing the costume of a feature.

Auditability isn’t one control — it’s the byproduct of several: immutable transcripts, an activity log of who accessed what, and version history for the prompts, disclosures, and routing rules that governed the agent’s behavior. When those exist, an audit is a query. When they don’t, it’s an archaeology dig.

• Store the full transcript for every interaction, including escalations and hand-offs, in tamper-evident form.
• Log access: who opened a record, when, and why.
• Version the agent’s configuration so you can answer “what were the rules on March 3rd?” not just “what are they now?”

5. The licensed-activity boundary: the line the agent never crosses

This is the row that sinks deployments. An AI agent can do an enormous amount of regulated work — intake, scheduling, status, document collection, first notice of loss — but there is a category of speech that legally requires a licensed human: diagnosis, a coverage determination, specific financial or investment advice, a legal opinion. The agent can gather everything up to that line. It must not step over it.

Consider the shape of it: an insurance agency’s AI can take the entire first notice of loss at 2 a.m. and open the record — it just can’t tell the caller whether they’re covered. That sentence belongs to a licensed producer. A medical practice’s agent can book the visit and collect symptoms; it can’t diagnose. The design work is deciding where the line sits in your vertical and making the agent hand off — warm, with full context — the instant a conversation approaches it.

1. Enumerate the licensed-only acts in your industry with counsel, in the agent’s own terms (“never state coverage,” “never diagnose,” “never recommend a specific security”).
2. Encode those as hard boundaries the agent cannot be talked out of, not as soft suggestions in a prompt.
3. Route to a named, licensed human on approach — early, with the transcript attached, so the customer never repeats themselves.
4. Sample transcripts weekly for near-misses and tighten the boundary from what you find.

Run the list before you go live

None of these five is exotic. What’s new is the speed and volume an AI agent brings — a control that was “good enough” when a front desk sent forty texts a day behaves very differently at four thousand. The teams that stay out of trouble treat consent, disclosure, retention, auditability, and the licensed-activity boundary as system properties they can demonstrate, not policies they wrote down once. If you can produce evidence for all five on demand, you’re ready.