Putting licensed providers behind your AI health product

Build a medical group or connect to one?

In all 50 states, prescribing belongs to licensed people, so your AI can’t write the prescription your users ask for. Connect to a licensed clinical layer by API: your AI takes the history, a provider licensed in the user’s state decides, and you pay per completed visit, about $33 a visit at 1,500 visits a month on a volume commitment.

Building the same layer yourself is a second company. Here’s the comparison.

Either way you keep the product, the conversation, the brand and the user. The rest of this post covers why software can’t cross the line, how the handoff works and what the bill looks like.

Why your AI can’t prescribe

Prescriptive authority comes from state licensure. Each state’s practice acts name the professions that may prescribe, and software isn’t on any list. Every e-prescription carries a licensed prescriber’s NPI. Controlled substances add identity-proofed, two-factor signing the prescriber completes personally.

People are testing the edges. Utah is running a supervised sandbox pilot for AI renewals of existing prescriptions, and federal bills have been introduced. A national product still has to work under today’s rule. The longer version is in Can AI prescribe medication?

Two more rules shape the build. Corporate practice of medicine doctrines keep clinical judgment with licensed providers, away from a lay company. Telehealth modality laws decide, state by state and treatment by treatment, whether async review is enough or a video visit is required. You can’t prompt around either. If you’re still weighing your own entity, see PC/MSO structures for venture-backed telehealth.

Write the handoff down before the prompts

Most teams leave the handoff to the model. Decide it in a table instead, and have a clinician review it.

Put three escalation rules in code, where a clever user can’t talk the model out of them:

1. Anything urgent (chest pain, suicidal thoughts, a severe reaction) exits to emergency guidance before a care request exists.
2. Clinical questions go to the care team. “Can I take this with my blood-pressure medication?” is a provider’s question.
3. When the infrastructure says no, because the state requires a live visit or the protocol excludes the user, the assistant explains the next step. It doesn’t retry with new wording.

Tell users they’re talking to AI and that a licensed provider decides. Our post on where AI stops covers escalation lanes in depth.

Connect by API now, MCP when it fits

• REST API. Your backend creates a patient, opens a care plan and submits an encounter with the intake your assistant collected. This is the production path today. Adding prescriptions to an existing app shows the calls.
• Webhooks. Review runs on human time, so outcomes arrive as events: encounter.in_review, encounter.escalated when the provider needs more, encounter.completed, then order.placed and onward. Payloads are thin; you fetch the resource, so PHI stays out of the webhook.
• MCP, in preview. An MCP server exposes the same actions as tools the model calls in the conversation. Lithos’s is in preview with design partners as of late 2026. The tools let an agent request care and follow it; deciding stays with the provider.

How much does the clinical layer cost?

Picture an assistant with 200,000 monthly active users. These inputs are illustrative; swap in yours.

• 1% ask for something only a clinician can give: 2,000 requests a month.
• 75% complete as visits after eligibility and intake drop-off: 1,500 visits.
• 85% async, 15% sync: 1,275 and 225.
• A volume commitment at the 1,000-visit band: a flat $30 async and $50 sync from month one.

Without a commitment, pay-as-you-go is graduated: visits 1 to 499 bill at $45 and $70, 500 to 999 at $40 and $60, and 1,000 and up at $30 and $50. At this mix that blends to about $42 a visit. A 50-visit pilot month costs 50 visits and nothing else.

Medication passes through at pharmacy cost, plus a small flat fee on mail-order orders. Add-ons are usage-based: benefits checks and prior authorization for insurance-covered branded prescriptions, and AI care agents.

Users pay you, through your own merchant account, whatever you set. Charge $79 a visit and the clinical layer is about 42% of visit revenue. Put care in a $49 monthly membership where a quarter of members have a visit each month, and it costs about $8 a member.

Add providers before users demand it daily

Waiting means bolting care on under pressure, with a handoff nobody designed. A sensible order:

1. Pick one category. Refills, weight care, sexual health and dermatology have well-defined protocols.
2. Design the handoff with a clinician.
3. Build against the sandbox. Sandbox routes drive encounters and orders through every status, escalations included.
4. Sign the BAA and settle protocols in parallel with the build.
5. Launch narrow, then widen. Track completed visits per request, escalation rate and time to decision.

Enterprise AI teams meet procurement early. Lithos signs a BAA at onboarding, runs on HIPAA-compliant infrastructure and has SOC 2 Type II in progress; see the security page and launching telehealth inside an enterprise.

Frequently asked questions