How to set up AI agents for a healthcare company.

1. Pick one job with a clock on it

The right first agent handles something time-sensitive, repetitive, and non-clinical: after-hours patient inquiries or applicant replies. Both have a measurable before (reply time in hours) and after (reply time in seconds), and both produce a booking. Do not start with clinical triage, billing, or "a chatbot for the website."

2. Write the scope before the prompt

Three lists, agreed with clinical and compliance leadership: what the agent may say (an approved answer bank: payors, hours, services, pay bands, schedules), what it must never say (clinical advice, diagnoses, unverified network status, anything about a specific patient's record), and what triggers a human handoff (symptom language, complaints, negotiation, anything outside the bank). The agent is only as safe as this document.

3. Handle the compliance prerequisites

• HIPAA: HIPAA-eligible infrastructure and a BAA for any workflow touching patient data. No PHI in ad platforms or analytics.
• TCPA and TCR: consent language on every form, campaign registration before the first text, opt-out handling.
• Lock-screen safety: message bodies carry logistics only.
• Disclosure: the agent identifies as your organization and does not claim to be human if asked.

The compliance checklist covers the marketing side of the same rules.

4. Connect the systems, not a widget

Phone number, CRM, scheduler. The agent should book real appointments and write real records, not collect a form for someone to re-key. If your scheduler has no API, start with the CRM and a shared calendar and upgrade later.

5. Run it in shadow mode

For one to two weeks the agent drafts replies that a human approves before sending. This is where the answer bank gets corrected and the escalation rules get tuned. Skip it and you launch with mistakes at scale.

6. Go live with a human in the loop

The agent sends; a named owner reviews every escalation daily and a sample of transcripts weekly. Widen scope only as the transcript review earns it: first replies, then booking, then reminders, then reactivation.

7. Measure to outcomes

Reply time (median and 90th percentile), inquiry-to-booked or applicant-to-interview rate, show rate, and escalation rate. If the escalation rate is very low, the scope is probably too loose. If it is very high, the answer bank is too thin.

Build, buy, or partner

Off-the-shelf chatbots are cheap and generic; they answer but do not act, and their handoff logic is not written for healthcare. Building in-house works if you have engineers who know both the API layer and HIPAA. Partnering with someone who has deployed healthcare agents gets you the scope document, the compliance setup, and the shadow-mode process without learning them the hard way. Provido builds agents on the Anthropic API and deploys them inside healthcare engagements.