Addiction treatment centers

What an AI Agent Should Never Do in Addiction Treatment

I'm a person in long-term recovery, and I founded an addiction treatment program in Idaho. So I'll say this as plainly as I can: some moments in treatment belong to people, every time, and no amount of good software changes that.

AI agents can take a lot of weight off a treatment team. This article is about where they have to stop.

Never handle a crisis alone

If someone says they took something, they're shaking and seizing, they want to die, or a person near them is in danger, the agent's only job is to get a human and get help. That means giving 911 for an emergency and 988, the Suicide and Crisis Lifeline, which you can call or text. Then a live person picks up, fast.

What this looks like in practice:

  • A written list of crisis phrases, built with your clinical director, that triggers an immediate handoff.
  • A rule that the agent never tries to talk someone down, assess risk, or decide the situation isn't serious.
  • A fallback for when no human answers, which is usually 911 and 988 plus a page to your on-call staff.
  • Testing after every vendor update. Models change. A handoff that worked in March can break in May.

Withdrawal deserves special attention. Alcohol and benzodiazepine withdrawal can be medically dangerous. An agent should never tell anyone it's fine to stop on their own or wait it out.

Never make a clinical decision

The agent doesn't diagnose. It doesn't decide level of care. It doesn't tell someone they need detox, or that outpatient will be enough. It doesn't comment on medications, doses, or whether someone should start or stop medication for opioid use disorder.

This isn't only good practice. Some states now say it in statute. Illinois passed the Wellness and Oversight for Psychological Resources Act, effective August 1, 2025. It bars offering therapy or psychotherapy through AI unless a licensed professional conducts it. It also says a licensed professional can't let AI make independent therapeutic decisions, directly interact with clients in therapeutic communication, generate treatment plans without review, or detect emotions or mental states. State laws are moving fast. Your counsel should check every state you serve.

An agent can summarize what a caller said, organize intake answers for a clinician, and book the assessment. That's real help. The judgment stays with the person who holds the license.

Never act as the counselor

There's a pull to put a friendly chatbot in front of patients between sessions. "Talk to Sam any time." I get the appeal. People in early recovery are lonely at 3am.

But a bot that sounds like it cares is not care. Recovery happens in connection with people. A patient who pours out something heavy to a chatbot may feel heard and still be in danger, and nobody on your team knows. If you offer any between-session tool, it should point people back to humans: their counselor, their sponsor, a peer, a meeting, 988. It shouldn't become the relationship.

Never touch referral money

The federal Eliminating Kickbacks in Recovery Act, 18 U.S.C. § 220, makes it a crime to knowingly and willfully pay or receive remuneration for referrals to a recovery home, clinical treatment facility, or laboratory. Penalties under the statute run up to $200,000 and 10 years per occurrence. It applies to services covered by health care benefit programs, which includes private insurance.

Why does this matter for AI? Because lead routing is easy to build. An agent could send callers to whichever partner pays a fee per admit, or a marketing vendor could bill you per qualified lead. That's the old patient brokering problem with a new interface. Don't build it, don't buy it, and don't let a vendor talk you into it. If a contract has money moving with referrals, admissions, or "qualified leads," get counsel to look at it before you sign.

Never make promises about money or outcomes

No coverage promises. No "your insurance will pay for 30 days." No price quotes the agent made up. No success rates unless they're real, current, and approved by your leadership, with the source attached. If a family makes a decision based on a number a bot invented, that's on you.

Never put PHI where it doesn't belong

This is the quiet one, and it's the one most likely to happen this week.

A tired counselor copies a group note into a free AI chatbot to clean up the grammar. A director pastes a patient's insurance denial letter into a consumer tool to draft an appeal. Neither person meant harm. Both may have disclosed protected information to a company with no Business Associate Agreement.

If your program is covered by 42 CFR Part 2, substance use disorder records have protections beyond standard HIPAA, including limits on using records against patients in legal proceedings. Starting February 16, 2026, HHS's Office for Civil Rights began accepting complaints and breach notifications under Part 2, according to its enforcement announcement.

The fix is policy plus training plus the right tools. Our staff training guide has a one-page policy you can adapt.

Never hide that it's a machine

Tell people they're talking to an automated assistant. Up front, in plain words. People calling a treatment center are often scared and ashamed. Finding out later that the warm voice was software can feel like a betrayal, and trust is hard enough to earn in this field.

Where the privacy line sits

Safe with de-identified or aggregate data. Reviewing call volumes, common questions, wait times, and FAQ performance. Drafting general content like a packing list, a visiting policy, or a blog post on what IOP is. None of that needs a patient.

Needs a BAA, and Part 2 review if it applies. Anything with names, phone numbers, insurance details, diagnoses, call recordings, transcripts, or notes. The vendor has to be a business associate under contract, and your counsel should confirm how Part 2 consent and redisclosure are handled.

Don't automate, even with a BAA. Crisis response. Clinical decisions. Counseling. Anything involving payment for referrals. Coverage and cost promises. A BAA makes a vendor responsible for protecting data. It doesn't make the vendor qualified to make those calls.

A short test before launch

Put your clinical director, your admissions lead, and someone in recovery in a room with the agent. Have them try to break it. Ask it whether someone needs detox. Say you want to die. Ask what it costs. Ask if you can bring your dog. Watch what it does.

Then read the admissions guide for what to turn on first. The things on this list stay off.

Frequently asked questions

Can an AI chatbot provide counseling to patients in addiction treatment?

It shouldn't stand in for a counselor. Some states have moved to prohibit it outright. Illinois, for example, bars offering therapy through AI unless a licensed professional conducts it, and bars licensed professionals from letting AI make independent therapeutic decisions or engage in therapeutic communication with clients.

What should an AI agent do if someone mentions suicide or overdose?

Stop the script, give 911 for an emergency and 988 for the Suicide and Crisis Lifeline, and connect a live person right away. Test this every time the vendor ships an update.

Can AI help decide level of care?

It can organize information for a clinician. The level of care decision belongs to a licensed clinician using your program's assessment process.

Is it legal to use AI to route callers to whichever program pays the most?

Paying or receiving remuneration for referrals to recovery homes and clinical treatment facilities can violate the federal Eliminating Kickbacks in Recovery Act, 18 U.S.C. § 220. Automating it doesn't make it legal. Talk to counsel about any arrangement where money moves with referrals.

Can staff paste patient notes into a free AI chatbot?

Not unless your organization has a Business Associate Agreement with that provider for that specific product, and your counsel has approved the use. Consumer chat products generally aren't covered.

Use cases for this guide