Can AI prescribe medication? How AI prescriptions actually work

The short answer

Prescriptive authority is a creature of state law. Each state’s practice acts say which professions may prescribe — physicians, nurse practitioners, physician assistants, and depending on the state some pharmacists, optometrists, and others — and within what scope. No state includes software on that list. Every electronic prescription identifies a licensed prescriber by NPI, and for controlled substances the DEA’s EPCS rules require identity-proofed credentials and two-factor authentication that the prescriber completes personally at signing. The one place the edge is being tested is Utah, where a state-supervised pilot lets an AI renew existing prescriptions — more on that below.

So an AI cannot prescribe. What it can do is almost everything around the decision — and that turns out to be most of the work.

What “prescribe using AI” means in real products

Three patterns sit behind the phrase, and only two of them are legitimate.

1. Clinician-in-the-loop prescribing. AI collects and structures the intake, screens it against protocol criteria, drafts a proposed plan, and queues it. A licensed clinician in the patient’s state reviews, decides, and e-signs — often asynchronously. This is how most modern telehealth programs work, and it is what agentic prescribing names precisely.
2. AI-assisted decision support inside the clinician’s tools. The model surfaces interactions, missing data, and protocol deviations while the clinician works. The clinician remains the decision-maker and can see the basis for every suggestion.
3. Software that issues prescriptions without a clinician decision. Whether by rubber-stamp review or none at all, this is outside every state’s rules and the pattern boards act on.

The workflow, step by step

1. Intent. The patient asks for care — in your app, or through an assistant acting for them.
2. Patient and consent. The agent creates or matches the patient; informed consent and identity proofing are captured before anything clinical.
3. Intake. A branched history: medications, allergies, contraindications, the vertical’s specific screen (a cardiac screen for ED, weight and comorbidities for GLP-1s).
4. Screening. Protocol eligibility, interaction checks, and state modality rules applied server-side. Ineligible or unlawful-as-requested cases never reach the queue as prescriptions; they route to a live visit or a decline.
5. Draft. The agent proposes a plan from the approved catalog — the starting dose, the follow-up interval — as a draft, not an order.
6. Decision. A licensed clinician in the patient’s state reviews the case and the draft, then approves, changes, or declines. This is the prescribing act.
7. Signature and transmission. The clinician e-signs; the prescription goes to a licensed pharmacy by e-prescribing.
8. Fulfillment and status. The pharmacy fills and ships; decision, shipment, and delivery events return to the agent over webhooks so it can tell the patient what happened.

The rules that shape it

What about Utah’s AI prescription pilot?

Utah is the first state to let an AI participate in prescription decisions, and it is worth being precise about what it did. In 2024 the state created an Office of Artificial Intelligence Policy with the power to sign “regulatory mitigation agreements” — a sandbox in which a named company gets temporary relief from specific rules while regulators watch. Under one such agreement, announced in January 2026 and running on a one-year term with an option to renew, the state’s licensing division agreed not to enforce unlicensed-practice rules against Doctronic so its AI can renew existing prescriptions for patients with chronic conditions such as hypertension, diabetes, and thyroid disease.

• Renewals only. The AI may renew a prescription a licensed physician already wrote. New prescriptions and any change in dose or frequency must be ordered by a physician.
• A bounded formulary. Roughly 200 maintenance medications. Controlled substances, ADHD medications, and injectables are excluded — so GLP-1 injectables, for instance, are out of scope.
• Physician review, for now. In the current phase a licensed physician must review each AI-recommended renewal before it goes to the pharmacy. A second phase, per medication group and subject to state approval, would let the AI send renewals directly to the pharmacist, who can escalate to a Utah-licensed physician.
• Limits and reporting. Patients get a limited number of AI renewals before they must see a Utah-licensed clinician; the company files monthly reports; the state publishes outcomes.

The early data, released in May 2026, is instructive for anyone building agentic workflows: the AI recommended renewal in about 72% of cases and escalated the rest to a physician for reasons like overdue labs, and where it recommended renewal the reviewing physician agreed about 91% of the time. That is a well-built collect-and-route system with a human decision at the end — the design this article describes — operating under an explicit regulatory exemption. The state’s own framing was “doctor, not device.” The medical licensing board objected anyway, and legal commentators have noted the agreement leaves open whether the AI is acting as a practitioner or as the physician’s delegate, since renewals go out under the named prescriber’s license.

What it means for a telehealth company: the direction is real, the mechanism is narrow. It is a pilot for one company, one use case, one state, on a one-year clock. Other states — Arizona, Texas, Wyoming among them — are building similar sandboxes, and the company involved has said it is in talks with several. Build like this is where the rules are heading, because it is; market and operate on the rules as they stand, which still put a licensed clinician on every prescription outside that sandbox.

And the federal bill?

A federal bill introduced in January 2025 would amend the Food, Drug, and Cosmetic Act so that AI or machine-learning technology could qualify as a practitioner licensed to prescribe, provided the FDA had authorized it and the state in which it operated permitted it. It names the two gates that would have to open: federal authorization of the specific technology, and a state granting prescriptive authority to software. At the time of writing the bill remains in committee with no vote taken, and no state has amended its practice act to grant that authority — Utah’s pilot works by suspending enforcement, not by changing who may prescribe. It is reasonable to expect the conversation to continue. It is not reasonable to build a program that depends on it.

Where AI adds the most value

• Complete intakes. A clinician reviewing a structured, branched history decides in a fraction of the time it takes to reconstruct one from free text.
• Pre-screening. Protocol criteria, contraindications, and interaction checks before review, so the queue holds decidable cases.
• Drafted plans. “Continue, step dose per protocol” as a proposal the clinician accepts or edits. See titration.
• Refills and renewals. Eligibility, follow-up collection, and the request itself — the highest-volume job in any program. Guide: Refill medication with AI.
• Follow-through. Pharmacy routing, tracking, and patient updates once the decision is signed.

Building it correctly

• Treat every AI output as a draft. The clinician’s action — approve, change, decline — is the only thing that creates an order.
• Enforce licensure, state modality rules, and the approved catalog in infrastructure, not in the prompt.
• Disclose AI involvement to patients where state law requires it, and design the disclosure to be true.
• Log what the model saw, what it proposed, who decided, and when. This is what a board, an auditor, or an acquirer asks for.
• Run controlled substances on a separate, stricter protocol with PDMP checks and EPCS signing.
• Keep clinical questions clinical: “can I take this with my blood-pressure medication” goes to a clinician, even when the model could plausibly answer. See Where AI stops.

Frequently asked questions