Why removing humans from care may undermine outcomes

The pitch deck version of digital health goes something like this: AI replaces the clinician, costs drop, access expands, outcomes improve, everyone wins. The pitch has been effective. Venture capital has poured billions into companies built around the premise that removing humans from the care loop is both possible and desirable. The premise has a problem, and the problem has a number.

Ruben Sandoval Davila, Co-founder and CTO Avena Health

Two percent. That is the share of active users remaining after three months on a clinical nutrition platform that, according to its own internal data, had previously retained 40% of patients for eleven months or longer. The platform is Avena Health. The person who built the automation, watched it fail, and then rebuilt the architecture around the failure is Ruben Sandoval Davila, the company’s co-founder and CTO. The data has not been independently audited, but the specificity of the claim, and the fact that Sandoval volunteers it as a story about his own mistake, lends it a credibility that rounder numbers would not.

The data point is uncomfortable for the industry because it comes from inside the house. Sandoval did not observe someone else’s automation experiment from a distance and publish a critique. He ran the experiment on his own product, with his own users, and the results were unambiguous. Full automation destroyed patient engagement. The AI performed its clinical functions correctly. The patients left.

“Healthcare doesn’t scale because providers are bottlenecked by time, not demand,” Sandoval said. “What we’re building shifts that constraint. The provider’s capacity isn’t limited by skill or willingness. It’s limited by the administrative overhead surrounding every patient interaction. Reduce that overhead intelligently, and the same provider can manage significantly more patients without degrading the quality of any individual interaction.”

The statement sounds like an argument for automation. It is, in fact, an argument for something more specific and harder to build. Sandoval’s position is that the bottleneck is real, the AI solution is partially correct, and the part that most companies get wrong is the assumption that efficiency and human involvement are inversely related. His data suggests they are not. The most efficient version of his platform, the one that produced a forty percent long-term retention rate, is the one that keeps human specialists in the loop at specific clinical touchpoints while automating everything else.

“Fully automated AI systems often struggle with retention,” he said. “The edge comes from combining hyper-personalization at the patient level with expert oversight.“

The insight is not that humans are better than AI at clinical care. It is that patients behave differently when a human being is involved in their care, even when the AI component is doing most of the work. The forty percent retention rate is not a measure of clinical quality. It is a measure of patient behavior, and patient behavior is what determines whether a digital health platform survives past its first year.

This matters because the digital health industry has a retention problem it does not talk about openly. Consumer health apps of all kinds, fitness trackers, mental health platforms, nutrition programs, chronic disease management tools, share a common lifecycle: rapid adoption, rapid abandonment. A 2023 analysis by the IQVIA Institute found that the average digital health app loses more than three quarters of its users within the first two weeks. Most of the venture-backed companies in the space report user acquisition numbers rather than retention numbers for a reason. The acquisition numbers look good. The retention numbers, in most cases, do not.

The industry’s response to this problem has been to double down on the engagement playbook borrowed from consumer tech: gamification, streak counters, push notification sequences, behavioral nudges. Sandoval’s response was different. He asked whether the problem was engagement at all, or whether it was something more fundamental about the absence of a human relationship in the care experience.

Sandoval’s architecture addresses this by treating the human specialist as a strategic asset within the system rather than as a cost center to be eliminated. The specialist does not do everything. The specialist does the things that produce the behavioral outcome the platform needs: continued engagement. The AI handles the rest.

“Agencies sell attention through marketing,” Sandoval said. “Experts deliver outcomes. The leverage comes from giving each expert the tools to own, control, and scale their growth.“

That framing explains why the platform is built around the specialist rather than around the patient or around the platform itself. The specialist is the retention mechanism. The AI is the efficiency mechanism. The architecture holds both in a relationship that Sandoval designed, tested against his own failure data, and refined across years of production deployment.

The implications extend beyond any single company. If the most important unsolved problem in digital health is not clinical capacity but patient engagement over time, then the companies optimizing for full automation are solving the wrong problem. They are building products that are efficient, scalable, and empty. The automation works. The users leave.

Sandoval is now preparing to test this thesis in the market where the AI hype is loudest and the retention problem is least discussed. Alva Health, the new platform his team is launching in the United States, is built on the same hybrid architecture that restored Avena’s retention rate after the automation failure. The U.S. digital health market is saturated with products that promise to remove the human from the loop. Sandoval is entering it with a product designed around the conviction that removing the human is precisely what breaks the loop.

The two percent figure is the kind of data point that an industry built on automation narratives would prefer not to confront. It suggests that the most important variable in digital health is not how much of the clinical workflow you can automate. It is whether patients stay. Sandoval has an answer to that question, and the answer required building the wrong thing first, measuring the consequences honestly, and designing an architecture that takes human attention as seriously as it takes algorithmic efficiency.

The industry may arrive at the same conclusion eventually. The question is how many retention curves have to collapse before it does.