Why Most Voice AI Claims in Healthcare Are Fiction (And What the Real Numbers Actually Show)

I built a mental health app once. Tracked mood, sent reminders, gave people a little interface to log how their bipolar disorder was affecting them day to day. Beautiful UX. Evidence-based prompts. The whole thing. I was proud of it.

Fifteen percent of patients used it weekly.

I don't mean 15% engagement was bad. I mean I had built something for sick people that 85% of them ignored entirely. That's not a product problem. That's a signal.

So I threw the app out and called them instead. With AI. And got 85% weekly engagement.

That number should make every healthcare operator stop and think about what they're actually building, what they're actually buying, and which vendor claims deserve scrutiny.

What Does the Evidence on AI Voice Agents Actually Show?

The honest answer is: it's mixed, and most vendor claims aren't peer-reviewed. Deepgram published a good piece in May 2026 cataloguing what the evidence actually supports before you build: booking lift claims of 30-50% from vendors, zero peer-reviewed backing. The real signal is that patients respond to proactive, conversational, phone-based outreach at rates that text portals and apps simply don't match.

What we know from the data: phone calls work. Patients pick up, especially when they're expecting a call from their care team. What isn't proven is that any voice AI system will automatically close that loop. The technology creates the capability. The engagement depends entirely on how it's deployed, what it says, and whether the patient trusts the voice they're hearing.

The state of the market in 2026 is instructive: 78% of enterprises have AI agent pilots. Only 14% have reached production scale. In healthcare, that drops to 8%. That gap isn't a technology problem. It's a deployment and clinical workflow problem that most vendors aren't honest about.

Why Are Conversion Rates from Voice AI So Variable?

The variance comes down to three things: timing, clinical relevance, and whether the patient knows to expect the call.

Platforms reporting 14% conversion rates on outbound recall campaigns (compared to 2-4% for postcards and generic emails) are doing something right in those three dimensions. But those numbers tell you about appointment reactivation, not clinical engagement. Those are different problems with different baselines.

At HANA, our 85% weekly engagement isn't a recall campaign metric. It's patients actively completing structured check-ins with a voice AI that knows their clinical context, asks relevant questions, and escalates when something's wrong. The use cases we've built for are post-discharge follow-up, chronic disease monitoring, and medication adherence. The call isn't just a reminder. It has clinical content.

That's the distinction most market reports miss: outbound dialing for scheduling is not the same as clinical outreach. They require different models, different prompts, different escalation logic.

Is Phone-Based AI Actually Better Than Apps and Text for Patient Engagement?

For most patient populations, yes. Substantially.

The research we've seen from our own deployments and from the broader literature consistently shows the same pattern: app-based and text-based interventions get high initial adoption among younger, commercially insured patients. They fall off sharply for older patients, patients managing multiple chronic conditions, and patients who are already overwhelmed. Those are also the patients you most need to reach.

Phone is the universal interface. The patient doesn't need to download anything, create an account, or remember a portal password. The call comes to them. That removes the largest single barrier to health technology adoption: the effort required to engage.

This is something I understood viscerally after years of building the wrong things. The Australian circus I worked with after my company collapsed had performers doing fire chains, not aerial silk. Why? Because fire chains are accessible. Anyone watching understands the stakes immediately. Aerial silk is technically harder, but the audience doesn't feel it the same way. The best patient engagement tools are fire chains. Direct. Felt. No translation required.

What Should Clinic Owners Ask Before Buying a Voice AI Platform?

Four questions that will tell you everything:

First, what's the clinical engagement rate, not the booking conversion rate? Those are different numbers, and vendors often conflate them intentionally.

Second, how does the system handle escalation? When a patient says something concerning, what happens in the next 60 minutes? If the answer is a flag appears in the dashboard for someone to check tomorrow, that's not a clinical tool.

Third, does it write back to your EHR, and how? A voice AI that produces PDFs to be manually reviewed defeats the whole point.

Fourth, what's the deployment timeline from contract to live calls? If the answer is longer than eight weeks, ask why. Most of that lag is the vendor's implementation burden, not yours.

The HANA integration and setup documentation is public. Read it before you read our sales materials. If you understand what we've built, the sales conversation is much shorter.

How Is HANA Different From Other Healthcare Voice AI Platforms?

HANA is fully open-source and self-hosted. No OpenAI dependency. No data leaving your infrastructure if you don't want it to. That matters for health systems that have learned the hard way what vendor lock-in costs.

We're also built by a clinical psychologist (hi, that's me) who has actually sat in rooms with patients, seen what disengagement looks like, and understands why engagement fails. The system was designed around clinical workflows first. The AI followed the clinical logic. Not the other way around.

After 1M+ patient interactions across 5 countries and 3 languages, zero critical adverse events, the data supports the design. 31:1 ROI across our clinic network. 85% weekly engagement from a patient population that averaged 15-20% before.

The market claims are loud right now. The evidence is quieter. But the evidence is what matters.

Key Takeaways

The voice AI market in healthcare is full of genuine capability and inflated claims sitting side by side. The gap between pilot and production is real, and it's not closing through product features alone. Clinical deployment is hard. Patient trust is earned over many interactions, not assumed. Engagement numbers only mean something when you specify what engagement means in your clinical context.

The platforms that will actually help patients are the ones that treat every call as a clinical interaction, not a marketing touchpoint. That means structured clinical content, real escalation paths, and outcomes you can measure. Everything else is noise.

FAQ

What engagement rate should I expect from a voice AI patient follow-up program?

Industry baseline for any form of post-discharge outreach is 15-20% weekly engagement. Well-deployed voice AI with clinical relevance, proactive outreach, and proper patient expectation-setting can reach 70-85%. The difference is almost entirely in the clinical design of the call content and the patient's expectation that the call is coming.

How long does it take to deploy a voice AI system for patient follow-up?

A properly engineered clinical voice AI platform should go live within four to eight weeks from contract signing. Longer timelines usually reflect poor documentation, complex custom integration requirements, or vendor understaffing. Ask for a reference site that went live in under six weeks.

Is it safe to use AI for clinical patient outreach?

With appropriate escalation protocols and human oversight design, yes. The key design requirement is that the AI must be able to escalate to a human within a clinically appropriate timeframe when a patient flags a concern. Pure automation without escalation pathways is not appropriate for clinical populations. HANA has completed 1M+ patient interactions with zero critical adverse events.

If you're a clinic owner or medical director evaluating voice AI for patient follow-up, I'm happy to show you the actual system. Book a 30-minute discovery call here.