Kinney Drugs handed its refill line to an AI voice assistant in May 2026 and named it Burt, after the chain’s founder. By late July, a Vermont customer told VTDigger that Burt calls nearly every day, mutters a drug name she cannot make out, and she says yes, because why else would the pharmacy be calling. She has four bottles of a medication she takes twice a week sitting at home [VTDigger, 2026]. Kinney Drugs is an advertiser with VTDigger, which the outlet discloses at the foot of the piece.
Her confused yes is a valid consent.
The Problem
Every state that permits telepharmacy regulates the same object: the pharmacist. Technician ratios, supervision, camera coverage, which tasks may happen at a remote site, how many prescriptions a day. Boards have spent a decade arguing over the span of one licensee’s attention, and Maine is still writing rules for a dispensing machine it already paid for.
The first conversation in that workflow now belongs to something nobody licenses. The voice agent takes the refill request, decides what enters the fill queue, and books the patient’s answer. It isn’t dispensing, so board rules don’t reach it. It isn’t clinical decision support, so FDA’s framework doesn’t reach it. And it isn’t a “health facility, clinic, physician’s office, or office of a group practice,” which is the exact list of covered entities in California’s AB 3030, codified at Health and Safety Code 1339.75, the first state law requiring that generative AI patient communications carry a disclaimer and a route to a human [Cal. Health & Safety Code 1339.75].
Read that statute against this case and the miss is precise. For audio communications it requires the disclaimer be spoken at the start and again at the end of the interaction, which is exactly the control that would have told a customer who was calling her about her medications. It binds a solo physician’s office. It does not bind the pharmacy filling that physician’s prescriptions.
The Insight
CMS built the one federal guardrail aimed at this outcome, and it wrote down the reason. Under the auto-ship policy announced in the 2014 Call Letter and reaffirmed since, network retail and mail-order pharmacies must obtain patient consent before each delivery the beneficiary did not personally initiate, including “refills prompted automatically to be filled at regular intervals” [CMS, 2014]. The agency’s stated purpose is not a signature. It is a check: contacting the beneficiary first “provides beneficiaries an opportunity to make sure that the pharmacy is delivering the correct drug (including strength, amount, and form) and if necessary, cancel or delay the order before it is billed and shipped” [CMS, 2014].
CMS even anticipated automation and drew a line at it. The same FAQ states the agency “does not support passive approaches for obtaining consent (for example, automated messages saying a member’s prescriptions will be automatically shipped unless the beneficiary takes action to stop them)” [CMS, 2014]. What it ruled out was an automated message the patient ignores. What arrived instead is an automated caller that extracts an active yes, which clears the rule as written and defeats every purpose behind it.
The economics are the mechanism. Getting consent used to cost something. A technician placed the call, waited through voicemail, tried again, and pulled an answer out of a real person. That labor was the governor. A pharmacy would not solicit consent it did not expect to convert, because asking was expensive. Drop the marginal cost of an outbound call to roughly zero and the rational operating point moves from calling the patients the fill history says are due to calling everyone, every cycle, and logging what they say. Each yes is documented, defensible, and billable. The compliance file gets cleaner. The patient gets four bottles.
“The safeguard was never the consent. It was the cost of asking for it, and that cost is now zero.”
The written record on who eats the risk already runs one direction. Kinney’s own terms state that Burt “may not be accurate, complete or up-to-date and may be misleading or contain errors and omissions.” McKesson, whose pharmacy system the pilot runs on, told investors that reliance on third-party AI “may expose us to risks that are outside of our control, including compliance gaps” [VTDigger, 2026]. The vendor disclaimed accuracy, the platform disclaimed control, and neither one is holding medication it did not need.
The success metric is the other tell. A Kinney director, quoted on McKesson’s own marketing page, said the tool is “reducing the calls to our pharmacy.” Customers described the downstream effect to VTDigger: confused by the AI, several now walk into the store instead, adding work at a counter that was already thin. Deflection is real and it is measured. The work moved to a channel nobody instrumented, which is how pharmacy AI keeps passing its own scorecard while the department feels worse.
Hospital pharmacy handled a comparable question in the open, and the contrast is the point. Nearly three-quarters of hospitals surveyed by ASHP use an EHR tool that automatically verifies certain medication orders without pharmacist review. ASHP convened a work group, published its findings in AJHP on July 31, 2025, and told members to take the practice to their state boards of pharmacy directly [ASHP, 2025]. Argue with the conclusions if you like. There was a public record, a professional body attached to it, and a named path to the regulator. The patient-facing version got a press release.
In Practice
Four layers sit between a patient and a filled prescription in a remote pharmacy model. The only one that talks to the patient is the only one with no regulator.
| Layer | Clear authority | What actually binds it |
|---|---|---|
| AI voice agent taking refill requests | Nobody named | Terms of service, surfaced after the fact |
| Technician at the remote site | State board of pharmacy | Ratios, supervision, task limits |
| Pharmacist verification | State board; DEA for controls | License, pharmacist of record |
| Dispensing automation | Board rules; FDA for the device | Accuracy standards, validation |
The Bottom Line
Vermont’s own answer is scheduled for later. Governor Phil Scott signed S.71, the Data Privacy and Online Surveillance Act, on June 16, 2026, eight days after vetoing it [EPIC, 2026]. It takes effect in two years, and one of the attorneys who testified on the bill told VTDigger he is not sure it would cover this situation even then [VTDigger, 2026]. Two years is roughly two dozen refill cycles per chronic patient, run by a system already documented to name the wrong drug.
The exposure lands where it always lands. Patients on several chronic medications, in towns with one pharmacy inside driving distance, cannot solve this by transferring, and CMS’s consent check was written specifically for them. The pharmacists absorbing the redirected traffic are carrying work the deflection metric says never happened. The beneficiaries are the vendors selling call-volume reduction and the operators booking the savings against a labor line.
What to watch is which regulator moves first, and whether it is a pharmacy regulator at all. A board of pharmacy could decide that an unsolicited AI refill solicitation is a pharmacy operation and therefore its business. A state attorney general could get there faster under ordinary consumer protection law, without waiting for anyone to define AI. Boards spent ten years settling how many remote sites one pharmacist may supervise. The unlicensed voice that now opens every one of those conversations was never on the agenda, and it is dialing today.