Agents in Recovery
The Behavioral Health AI Agent Readiness Checklist
Answer every line before an AI agent touches your phones, your intake, or your records. A "no" doesn't mean never. It means not yet.
1. The job
- We wrote one paragraph describing the job: what the agent does, what information it touches, and when it hands off to a person.
- One named person owns the agent, reads its conversations every week, and has the authority to shut it off.
- We know what success looks like (for example, after-hours calls answered and called back by 9am) and how we'll measure it.
- We have a written plan for turning it off, and staff know how to cover if we do.
2. The data
- We listed every type of information the agent will see, store, or send.
- We checked whether the job can be done with de-identified or aggregate data instead. (Removing names isn't de-identification. HIPAA's Safe Harbor method lists 18 identifier types.)
- The agent gets only the minimum information it needs, not the whole chart.
- We know where recordings, transcripts, and logs are stored, for how long, and who can see them.
- The vendor told us in writing whether our data is used to train or improve its models.
3. Law and contracts (bring these to counsel)
- Counsel has confirmed whether we're a HIPAA covered entity or business associate for this use.
- We have a signed Business Associate Agreement for the exact product and plan staff will use.
- We have the vendor's list of subcontractors that handle our data (phone, transcription, cloud, AI model provider).
- If we're a federally assisted SUD program, counsel reviewed how the tool handles 42 CFR Part 2 consent and redisclosure.
- Counsel checked state AI and recording laws for every state we serve (some states restrict AI in therapy).
- Any AI-voiced calls or texts have consent language reviewed under the TCPA.
- Pages with chat widgets or intake forms have no third-party ad or tracking pixels that could capture what people type.
4. Safety
- Our clinical director approved a list of crisis phrases that trigger an immediate handoff to a person.
- We tested the crisis route: the agent gives 911 for emergencies and 988 for the Suicide and Crisis Lifeline, and a human gets on the line fast.
- There's a fallback when no human answers, and it's been tested at 2am.
- The agent refuses clinical questions (diagnosis, level of care, medication) and routes them to a person.
- The agent tells people up front that it's automated.
- The agent never promises coverage, cost, or outcomes.
- We rerun our test conversations after every vendor update and keep the results.
5. Money and referrals
- Vendor pricing is flat or usage based (seats, minutes, conversations).
- No payment to or from anyone varies with leads, referrals, admissions, or placements. (See 18 U.S.C. ยง 220, the Eliminating Kickbacks in Recovery Act.)
- The agent never routes callers to another organization in exchange for anything of value.
6. People
- Staff have a one-page AI policy: what's allowed, what never goes into a tool, and what no tool does.
- Everyone who touches the agent or sensitive data is trained, including house managers, peers, front desk, and night staff.
- Staff know how to report a mistake the same day without fear of punishment.
- We review a sample of real transcripts as a team at least monthly.
7. Launch
- We start small (for example, after hours only) with a person reviewing every conversation for the first few weeks.
- We wrote at least 30 realistic test conversations, with no real patient details, and the agent passed them.
- The agent is in our HIPAA security risk analysis.
- We set a date to decide whether to expand, fix, or stop.