We're Building Health AI Like It's SaaS. That's the Problem.

There's a number that should embarrass the entire digital health industry: 3.3%.

That's the percentage of health app users still active by day 30, drawn from a peer-reviewed analysis of unguided mental health apps in real-world conditions (Baumel et al., 2019, JMIR). Not day 365. Day 30. Billions of dollars raised, thousands of engineers hired, millions of downloads celebrated in press releases, and within a single month, over 96% of users have quietly walked away.

We keep calling this a retention problem. It isn't. It's a founder problem. A mindset problem. And until the people building health AI are honest about where it comes from, the next billion dollars will produce the same result.

We Built for Return Visits

Most health AI founders come from software, finance, or consulting. They know how to find large markets, assemble teams, and iterate toward product-market fit. What many have never done is spend real time inside the healthcare system as a patient, a caregiver, or a clinician. So they do what they know. They frame the problem as a distribution problem: care is inaccessible, technology can scale it. They hire product teams trained to chase engagement. They optimize for the metrics their investors recognize.

The result is a SaaS product wearing a healthcare badge.

SaaS (Software as a Service, the subscription model that has dominated the last two decades) works because engagement and value are aligned. A CRM becomes more useful the more data you feed it. A project management tool gets stickier the more your team lives inside it. Use equals improvement.

Health is the opposite. A great diabetes management platform should be working toward a patient who eventually needs it less. A mental health tool succeeds when someone is genuinely equipped to handle difficulty, not when they open the app every morning. In health, the goal is to make yourself unnecessary.

But that's not what we built. Consider how a typical digital health startup measures success: daily active users, monthly retention, push notification open rates, session length. These metrics say nothing about whether anyone got healthier. They measure return visits. In SaaS that's revenue. In health it can be a warning sign.

Headspace is an honest example. At its peak it was commercially successful by every SaaS measure: over 70 million downloads, top App Store rankings, strong subscriber retention. Yet a systematic review of randomized controlled trials evaluating Headspace and Calm found outcomes "mixed" and studies "generally underpowered to detect small or medium effect sizes," with conflicts of interest common in the research base (Linardon et al., 2022, JMIR mHealth). That doesn't make Headspace a bad product. It makes it a precise illustration of the gap between commercial success and clinical impact, the product of a model that isn't designed to tell the difference.

What Health Actually Requires

Health doesn't fit the SaaS session model because it isn't episodic. It is a continuous story shaped by decisions made years before any symptom appears. A 45-year-old's cardiovascular risk was built in their 30s. The tools we are building don't know that. They know your last session.

But the deeper problem isn't memory. It's context. The most predictive signals for a health crisis aren't in your bloodwork, they're in your life. Job loss increases cardiovascular event risk significantly in the months that follow (Dupre et al., 2012, Archives of Internal Medicine). Bereavement predicts measurable immune suppression (Buckley et al., 2012, Brain, Behavior, and Immunity). These are clinically documented relationships between life events and biological deterioration. No health AI today ingests that. A continuous glucose monitor tells you your blood sugar is spiking. It doesn't know you just went through a divorce and haven't slept in three weeks. A cardiologist would ask. The app doesn't.

Build and Measure Differently

The constructive question is what a genuinely different approach looks like: in product terms and in the numbers used to judge success.

Observe, don't ask. The best health intervention is one the patient doesn't have to remember. Current tools ask too much: daily check-ins, manual logging, active decisions. The next generation needs to observe rather than prompt. A 2026 COPD trial tracked 220 patients using wristbands that predicted severe exacerbations with 84.8% accuracy, detecting events 4.4 days before clinical confirmation with zero active input required (eMEUSE-SANTÉ, PLOS Digital Health, 2026). The technology exists. The product philosophy doesn't yet.

Design for graduation, not retention. The most radical decision a digital health company could make is defining success as the user needing them less. Build an explicit pathway from high-touch AI support to guided self-management to occasional check-in. It is a completely different product architecture. It is also the only one whose incentives are genuinely aligned with the patient's, and the only one a serious payer should trust.

Build a health narrative, not another app. Your health story is fragmented across a primary care EHR, a specialist system, a wearable, a pharmacy platform, and three wellness apps that don't talk to each other. The missing product isn't another app. It is a persistent intelligence layer that accumulates your story across time and surfaces it when it matters. Not a record. A reasoned narrative.

Then measure accordingly:

1. Behavior Durability Score instead of DAU. Not how often someone opens the app, but whether a target health behavior persists 90 days after initial adoption: medication adherence, consistent sleep, daily movement. The behavior is the product. The app is the delivery mechanism.
2. Clinical Graduation Rate instead of Churn Rate. What percentage of users reach a defined health milestone and transition to lower-intensity support? Only one in three healthcare organizations currently tracks outcome-linked metrics like chronic-condition follow-up (KLAS Digital Health Most Wired, 2025). A product that can show this number has a genuine commercial edge.
3. Avoided Cost instead of LTV. The financial value of a health AI product isn't what patients pay per month. It's what the system doesn't spend. An avoided heart failure hospitalization is worth $15,000 to $30,000. A prevented ER visit is worth approximately $2,000. These are the real unit economics of health, and they are the language payers actually speak.

Why This Is Now a Commercial Imperative

The graveyard of the last digital health funding cycle is full of companies that optimized for engagement, raised on DAUs, and couldn't answer the only question that ultimately matters: did anyone actually get healthier?

The next shakeout won't be about technology. It will be about legitimacy.

Payers have been burned too many times by tools that promised outcomes and delivered dashboards. Health systems are developing real procurement rigor around clinical evidence. Regulators are moving fast: in 2026 alone, 43 states introduced over 240 healthcare AI bills (Manatt Health AI Policy Tracker, 2026). The era of building fast and asking permission later is closing.

The companies that reframe their metrics now aren't being idealistic. They're being strategic. The buyers of the next generation of health AI will demand proof of outcomes, not proof of engagement. The product that shows it made users measurably healthier wins the contract. The product with strong DAUs and 4.8 stars in the App Store gets politely passed over.

The question was never whether AI could transform healthcare. It clearly can. What's missing isn't technology or capital. It's the willingness to define success the way a patient would: not by how often they opened the app, but by whether their life got measurably better. The companies that make that shift won't just build better products. They'll build the ones that last.
‍

Luca Caruso is a digital health entrepreneur and operator with 17 years of experience at the intersection of healthcare, technology, and capital. Born in Italy, shaped by Europe and the US, he has spent his career helping digital health innovation actually work in the real world. He has reviewed over 300 health tech companies and partnered with international founders expanding across the US and European markets. Today he operates through Health Outpost, helping MedTech and HealthTech companies navigate US- market entry, and So-Talented, focused on building leadership teams where execution really matters. He also organizes Health2Tech in Miami, teaches digital health entrepreneurship at ESADE, EADA, and UN-affiliated programs, and previously founded Future Health Club in Barcelona.

‍

Sources

Baumel et al. (2019), JMIR mHealth and uHealth — real-world retention analysis of unguided mental health apps

Linardon et al. (2022), JMIR mHealth — systematic review of mindfulness app RCTs

Dupre et al. (2012), Archives of Internal Medicine — job loss and cardiovascular risk

Buckley et al. (2012), Brain, Behavior, and Immunity — bereavement and immune suppression

eMEUSE-SANTE trial, PLOS Digital Health (2026) — COPD wearable monitoring

KLAS Digital Health Most Wired National Trends (2025)

Manatt Health AI Policy Tracker (2026)