Avena rebuilt its AI nutrition platform around humans after retention fell to 2%

Avena Health co-founder and CTO Ruben Sandoval Davila told The Next Web in May 2026 that he ran the automation experiment on his own product: after the company fully automated care on its clinical nutrition platform, only 2% of active users remained after three months — against 40% of patients retained for eleven months or longer under its human-involving model, by the company's own internal data, which has not been independently audited.

His conclusion: efficiency and human involvement are not inversely related. The version of the platform that produced the 40% long-term retention keeps human specialists in the loop at specific clinical touchpoints while automating everything else. "The specialist is the retention mechanism. The AI is the efficiency mechanism." The industry context: a 2023 IQVIA Institute analysis found the average digital health app loses more than three-quarters of its users within two weeks, and most venture-backed companies in the space report acquisition numbers rather than retention numbers.

Sandoval is now preparing to launch Alva Health in the United States on the same hybrid architecture, entering a market saturated with products that promise to remove the human from the care loop.

Fully automated AI systems struggled with retention even when the clinical functions worked correctly.

Patient behavior — continued engagement — decides whether a digital health platform survives past its first year.

The binding constraint in healthcare is provider time and the administrative overhead around each interaction, not clinical skill.

Automate the work, not the relationship: measure what keeps users present, and design the human touchpoints around that measurement rather than around cost.

Avena restored its roughly 40% long-term retention on the hybrid architecture, according to Sandoval. The team is taking the same design to the US market under the Alva Health name. The retention figures remain self-reported and unaudited, as the article itself flags.