How we keep AI agents safe in a clinic: seven rules

Before an AI agent talks to your patients, you should be able to answer three questions about it. What is it allowed to do? What does it do when it isn't sure? How would you find out if it went wrong?

If a vendor can't answer those in plain words, the problem isn't the AI. It's that nobody wrote the rules down.

Here are ours. They apply to every role on our agent roster, from the Receptionist to the Prior Authorization Specialist, and none of them depends on the AI having a good day.

1. Every agent acts from a short list

An agent can only take the actions on its list: check open times, book an approved service, send a confirmation, send a deposit link, email your desk, put a caller through. If an action isn't on the list, it can't happen, however the conversation goes.

You decide what's on the list for your clinic. We bring those decisions to you as a one-screen table of yes-or-no answers, prefilled with the sentence the agent would say. A few sample rows:

Sample rows. Every clinic fills in its own.

Where a rule matters, software enforces it, not the AI. Your medication list, checked by code, decides which refill requests go to the desk for approval. Your hours, read on every call, decide whether a caller is put through. Your deposit rule sets the amount. The AI can't talk its way past any of them.

2. Clinical lines are never crossed

Our agents never prescribe, never suggest a dose, never interpret a result, never tell a patient a symptom is normal, and never read the chart aloud or confirm to anyone else that someone is a patient.

A problem after a treatment gets its own path. When a patient calls or texts about something that sounds wrong, your team gets an urgent alert right away, and it's never treated as a sales lead. On texts, that check runs before the AI sees the message, so an AI outage can't swallow it. After every call, the transcript is checked again for the same warning signs, in case something slipped past during the call. The patient is told the team has it only after the alert has actually gone out. Words that suggest something severe get a plain instruction to call 911.

At our first clinic, the owner approved the list of 68 phrases that count as urgent before it went live. The first real report came in on September 30: the patient was put through to the desk, and the urgent email was delivered.

3. Testing mode comes first

Before a clinic goes live, its texts can only reach approved test numbers. The one exception is a patient who has just called or booked: for the next hour they get the texts about their own request, like a booking confirmation. Reminders, follow-ups and win-back texts don't reach your patient list until you say go.

When we need to test something new, we add a tester to the approved list. We never switch the protection off to test.

4. Every call is reviewed, and every change is versioned

Every morning, the previous day's calls are checked against your rules. The review looks for a price you haven't approved, a guess at how many units someone needs, a promise to text a link that never went out, a transfer that didn't really happen, and a call that didn't end cleanly. When it finds one, we fix the cause.

Every change to your clinic's settings is saved as a new numbered version, so any change can be undone by putting the earlier version back. Your receptionist's playbook is versioned too, and each clinic is pinned to its own version, so one clinic's change doesn't move another's. Every day, a separate check compares what's running on your phone line with what we built, and flags any difference. And the AI model behind your line is pinned: a newer model doesn't reach your patients just because it's newer.

5. "Done" means confirmed

A system saying "request accepted" isn't the same as the thing happening, and we treat them differently. A booking counts when it shows up in your EHR. If the chart rejects a booking or never confirms it, we get a text within minutes. A desk request counts when the email is delivered. Payments are read back from your register.

From mid-September to October 1, our agents sent the desk at our first clinic 160 emails, and all 160 were delivered.

6. A quiet day is written down, not left blank

A job that did nothing and a job that never ran look the same unless someone writes down the difference. So our scheduled jobs, including the daily call review, the booking watch, the follow-up texts and the call-back check, record each run, empty ones included, with the reason they were empty: nothing was due, the patient had already booked, the caller wasn't a quick hang-up.

When a job goes quiet, the record says why, so a slow day never gets mistaken for a broken one.

7. HIPAA compliant, with a signed BAA

Radarrr is HIPAA compliant, and we sign a BAA with every clinic before go-live. Call recordings are private and can only be reached through an authenticated request. Patient texting is carrier-approved: the carrier reviewed the exact messages patients receive. A patient who texts STOP stops getting every kind of message from us, and marketing texts respect quiet hours.

Questions to ask any AI vendor

Use these with us or with anyone else.

1. What can the agent do in my systems? Can I see it as a list?
2. What does it do when it isn't sure?
3. What will it never say to a patient?
4. Who decides what it's allowed to say, and how do I change that?
5. How do you test before it texts my patients?
6. How do you know a booking actually landed in my EHR?
7. If something breaks at 2 a.m., who finds out, and how?
8. Will you sign a BAA?

If the answers come back specific, you're talking to someone who has run this in a real clinic.

To hear an agent working under these rules, try Gigi at radarrr.io/clinic/try.

How we counted: desk emails from the email service's delivery records, September 14 to October 1, 2026. Urgent phrases from the clinic's approved list.