AI at the Front Door of Care
Guest: Alex Cohen | Released 11/05/26
Season 2 of STAT AI opens at the front door of care: the phone line. Alex Cohen, co-founder and CEO of Hello Patient, joins DJ and Ethan to talk about conversational AI agents that answer patient calls, book appointments, handle billing questions, and follow up after visits. The conversation keeps returning to one question: what can an agent take off a team's plate, and where does it need to stop and hand the work to a person?
Cohen's path into the problem started at Carbon Health, where he was asked to fix a call center that was missing a large share of incoming calls during the peak of COVID. He went on to lead the company's patient-facing products, from online booking to the paper statement in the mail. Missed calls, slow callbacks, and post-visit follow-up kept coming up, and he describes them as problems nearly every outpatient group shares. At the time, automating even a simple reschedule over the phone was close to impossible. Today, Hello Patient builds the orchestration layer that lets an agent hold a full conversation, read context from the practice's systems, and complete the task end to end.
He walks through what that looks like for a parent calling an urgent care after a playground injury: confirming the clinic treats children, checking insurance, finding a location with X-ray on site, negotiating an appointment time, and creating the child's record linked to the parent. He is direct that none of this works out of the box. Implementation takes weeks of discovery with operations leaders and front-desk staff to capture what he calls the "dark matter" of a clinic: the local rules, spreadsheets, and habits that never made it into a manual.
Safety and escalation run through the interview. Cohen describes red-flag screening that looks for a possible emergency by meaning rather than keywords, practice-specific guardrails, and slower reasoning models that review conversations afterward and flag calls for a human to revisit. Agents are told to be conservative and transfer when in doubt. He is also clear that no company can promise a perfect record, and that practices should weigh a small margin of error against the calls that currently go unanswered. Calls that are clinical or outside an agent's scope, such as missing lab results or post-surgical concerns, go to the clinic team.
The group also discusses what happens to front-desk staff, integration with electronic health records, and patient satisfaction. Cohen reports that groups more often stop backfilling roles or move staff to higher-value work than cut positions, and that front-desk answer rates rise once routine calls are handled. EHR access remains the limiting factor: some vendors offer open APIs, while others resist third-party access. He closes with a look ahead at agent-to-agent conversations, where a patient's own AI assistant and a clinic's agent resolve a request together.
The hosts open the episode with Epic's October 2026 announcement of Connectors, which let Epic's AI agents reach third-party applications using the Model Context Protocol. They separate it from Epic's public MCP server for developer documentation and lay out the questions they would ask before turning on any healthcare agent: what can it read and change, can its actions be audited, and how does it recover when something fails?
What we cover
📞 Why missed calls and slow callbacks are a shared outpatient problem
🤖 How a voice agent books, reschedules, and handles billing calls end to end
🧭 Discovery, practice-specific protocols, and the "dark matter" of clinic workflows
🚩 Red-flag screening, escalation to humans, and post-call review
👩💼 What changes for front-desk teams when routine calls are automated
🔌 EHR integration, write-back, and the limits vendors place on access
🔗 Epic Connectors, MCP, and questions to ask before enabling an agent
Resources discussed
Epic launches Connectors to link its AI tools with third-party apps, Fierce Healthcare coverage of the October 21, 2026 announcement discussed in the hosts' opening segment. The conversation reflects the information available at recording.
