The Machine That Remembered Her

The front door of American healthcare is a telephone, and it does not work. Industry estimates for how many calls to a practice go unanswered run from roughly a quarter to over 40 percent. Roughly half of the people calling wanted to book a visit. Practices have thrown IVR trees, answering services and offshore call centers at this for two decades, and the number has not moved much.

Raihan Faroqui, an internal medicine physician who leads partnerships at Confido Health, has a theory about why the software never landed, and it is not the one vendors usually offer.

The migration nobody wanted

Healthcare, he argues, has a system of record problem. A practice already runs five or six of them: the EHR, the practice management system, revenue cycle, the phone system. Historically, adopting anything new meant migrating off one of them, which is why so much good software died in procurement rather than in production.

Agents skip that. “What we’ve actually needed is a system of action that is able to plug into that system of record that exists,” he told us. Confido changes none of them. Its agents weave through what is already there.

That reframing explains the adoption curve better than any claim about model quality. It also sets up the more interesting finding.

The product might not be the agent

Confido sends what it calls forward deployed engineers, a model it takes openly from Palantir, to sit physically alongside receptionists and billers and watch them work. What those engineers find is not a technical specification. It is undocumented practice.

“A lot of healthcare operations is just a bunch of Post-it notes on the computer,” Faroqui said. Dr. Smith only takes new patients Monday, Wednesday and Friday between three and five. That rule is written nowhere. The work is extracting it, writing an SOP, and training an agent on it, because agents have to be trained on data and logic.

Leon put the obvious question to him on tape: is the product the agent, or is it the AI literacy you are delivering to the customer’s team? Faroqui did not resist the reframe. It is the same finding this show has arrived at repeatedly. The model is the easy part.

There is real engineering too. They orchestrate across the record systems a practice already runs, train small language models tuned per specialty on their own accumulated voice data, and pair every agent with a monitoring layer. Their definition of failure is unusually precise and unusually modest: not harm, but acting outside protocol. That rate ran 5 to 10 percent at launch and sits at 1 to 2 percent now.

And then the part that is harder

One customer serves a largely Medicaid, largely Spanish-speaking population across the Bronx and Westchester. They told Confido the stock Spanish was unusable, because Dominican Spanish is not the Spanish of someone raised in the Bronx. So the team recorded the local dialect and built the agent to match it.

An outbound reminder reached a woman overdue for a mammogram. She began to cry. “I can’t believe you remembered me,” she said. A family member had died recently. She had been feeling low.

Nobody had called her. Then something did.

The agents do not identify themselves as AI, and by Faroqui’s account about half of patients do not know. That is the practice’s decision, not the vendor’s, so we set the ethics aside and asked the practical version: would that patient’s trust survive finding out? His answer was that she would have cried anyway, because the practice still remembered her, just through a different modality.

It is a strong argument. It is also the whole tension, and the episode does not pretend to resolve it.

Listen to the full conversation.

You Might Also Enjoy

• S1E49 with Christine Dymek — The AI literacy framework this episode keeps bumping into, from someone who built it deliberately rather than discovering it in the field.
• S1E36 with David Hidalgo-Gato — The other side of the same unglamorous coin, automating charge capture where the money quietly leaks out.
• S1E26 with Giovanni Donatelli — Language access as a clinical problem rather than a courtesy, which is exactly what the Spanish dialect story turns out to be.