AI

Sierra says an AI agent doubled mammogram bookings. The next measure is completed screenings.

Clay Bavor told Stephanie Davis how Sierra defines success for AI agents that book visits and answer benefits questions, and why healthcare organizations still have to track what happens afterward.

Author
HDM Staff
Save
Share
Stephanie Davis with an open notebook speaks with Clay Bavor on the HLTH Main Stage
Stephanie Davis interviews Sierra co-founder Clay Bavor in Las Vegas. Photo credit: HLTH USA 2026.

    Sierra co-founder Clay Bavor said one of the company’s AI agents doubled mammogram scheduling in an outreach effort. The task involved more than getting a patient to pick up. In the example he described, the agent contacted someone overdue for screening, asked about their availability, checked three nearby clinics and followed up with an appointment and reminders. The increase Bavor cited was in scheduling. He did not identify the provider or say how many people ultimately had the screening.

    That distinction matters because Sierra says it charges customers when its agents get something done. Bavor explained the company’s approach in a Nov. 17 Main Stage interview with Stephanie Davis, co-host of Rounds, in Las Vegas. He cited successful appointments and referrals, along with resolution of claims or benefits questions, as examples of results the company can bill for. “If and only if that happens, is there a charge,” he said. A health system using that model has to decide what a finished job looks like, particularly when a patient’s need takes days or weeks to resolve.

    Davis asked how software that first earned its place in retail could handle healthcare. Bavor said Sierra drew on work in retail and banking where agents had to deal with bookings, changes and other actions that require more than an answer. For healthcare, he said, the company also recruited people familiar with the systems and the ways payers and providers use them. The technical question is whether the software can act on available data. The operational question is whether a clinic or health plan is ready to let it take the next step.

    The result depends on what the contract counts

    An appointment put on the calendar is easy to count. A patient can still discover that the slot is at an inconvenient location, falls outside coverage or requires an order another office has yet to send. Treating every booking as a success can put the agent and the patient on different paths. For this kind of outreach, a provider could track the original list of eligible patients, how many were reached, how many accepted a slot and how many completed the exam. Those measures answer different questions about the workflow.

    The same distinction applies when a member asks about a claim. An agent might accurately explain that the claim is pending. If the caller needs to learn whether additional documents are required, the answer about status may save a minute while leaving the problem in the same place. Bavor argued that organizations should make their online and phone services capable enough that “you’re not project managing your own healthcare.” The example highlights what a buyer would have to specify. The relevant result could be a correct explanation, a completed submission, or a referral to a person who can act on the case.

    Bavor said Cigna went from its first conversation with Sierra to a live deployment in 56 days. Sierra had previously described an eight-week implementation and reported an 80% reduction in the time required to verify a patient’s identity. Those are vendor-reported measures of implementation and call handling. A separate Cigna customer account from Sierra describes a broader effort to make information and follow-up easier to navigate. The deployment figures do not measure whether a claim was resolved or whether a member received care sooner.

    An agent may pick up where the last call stopped

    Davis pressed Bavor on what happens after a single encounter. A patient seeking specialty care can face a referral, questions about authorization, a wait for available appointments and multiple follow-ups. Bavor argued that agents have become capable of tracking some tasks over days or weeks. He pictured the mammogram agent following through on outreach and allowing a patient to ask for a call back later, including at a particular time. A referral or other complex process adds more dependencies and more opportunities for work to stall.

    Bavor described one difficulty that has little to do with how natural a voice sounds. Primary care and specialist offices can use different appointment codes or categories. The person arranging the visit may have to know how both offices work to match a referral to an appropriate slot. Automating that step requires reliable access to the right information and agreement on what each scheduling option means. A fast call does little for a patient who reaches the wrong office and starts again.

    Davis also asked what would happen if an organization’s agent called an agent acting for the patient. Bavor said a Sierra agent on the provider side had already called one of its agents on the payer side. He expects that systems may eventually be able to discover how to interact directly through shared technical standards. That was a prediction about how organizations might exchange requests. A production workflow would still need to establish who can act for the patient, what each side is allowed to share and how a person can see or correct the result.

    Patient access still depends on available care

    Davis returned to the problem of capacity. If the patient reaches a responsive service and there is no clinician or appointment available, an automated exchange cannot create one. Bavor said Sierra sees opportunities to help with administrative steps around referrals, follow-up after procedures and routine outreach. He mentioned the recent hiring of Dr. Mitesh Patel, previously a clinical transformation executive at Ascension. Sierra’s hiring announcement describes the company’s work with clinical teams. Bavor said direct medical advice is outside the scope he expects Sierra to pursue.

    The work near that boundary raises a practical question. A patient may begin a call asking to change an appointment and then describe something that needs a clinician’s attention. The organization has to decide who takes over and what information travels with the handoff. For ongoing outreach, it also has to decide how many times to call and when to stop. Those decisions belong in the workflow before any system is evaluated on how many tasks it closed.

    Sierra’s screening example offers a starting point for that evaluation. A clinic can measure whether it reached more of the people it intended to reach and whether the appointments fit patients’ schedules. It can then follow the bookings through to completed exams. Bavor gave the first indication that the scheduling step moved. For the practice and the people it contacted, the result that matters comes after the reminder, when the patient can get the screening the outreach was meant to arrange.