We Built a Mental Health App. Nobody Used It. Then We Started Calling.

I built a mental health app for patients with bipolar disorder. Good product, honestly. Clean interface, peer support, mood tracking with actual clinical logic behind it. We were proud of it. We launched it properly, got it into the hands of the care team, trained the staff. Fifteen percent of patients engaged beyond week two.

I threw it out.

Not because the product was bad. Because I'd been solving the wrong problem. Patients with serious mental illness don't open apps when they're in crisis. They don't open apps when they're stable either. (Turns out, neither do most healthy people.) The app required the patient to remember, to initiate, to feel well enough to engage. It assumed motivation. And motivation is exactly what illness steals first.

So we started calling them instead. With AI. Eighty-five percent weekly engagement.

That number is real. It's not a cherry-picked pilot. It's across more than a million patient interactions, in five countries, three languages, real clinical environments with real sick people. I want to talk about what it actually means, because I think the healthcare industry is fundamentally misreading it.

Why Do Patients Ignore Healthcare Apps but Answer Phone Calls?

Answering a call requires almost no activation energy. The phone rings and you pick up. That's it. An app requires you to remember it exists, find it on your phone, open it, navigate to the relevant feature, and feel like engaging. That's five or six friction points for someone who might be post-surgery, confused about their medication, or just old.

I thought about this a lot during a circus tour I did through the Australian desert. (I know. Bear with me.) The performers drawing the biggest crowds weren't the ones thirty feet up on aerial silk. They were the ones doing fire chains at eye level. Accessible. Immediate. You walked past and you were already watching before you consciously decided to watch. That's what a phone call is. It comes to you. You don't have to go anywhere.

Voice AI patient engagement works for the same reason. Not because the technology is more sophisticated than an app. Because it meets patients exactly where they already are, without asking them to do anything first.

What Engagement Rate Should You Expect From AI Patient Follow-Up?

A well-implemented AI patient follow-up system should achieve 70-90% weekly engagement rates. The industry baseline for digital tools like apps and patient portals sits at 15-20%. That gap isn't random and it isn't a design problem. It's structural.

Passive channels — apps, portals, SMS links — rely on patient initiation. The patient has to come to the tool. Outbound voice flips that completely. The care team reaches out. The patient responds. That inversion is everything when you're dealing with people who are unwell, overwhelmed, or simply not oriented around managing their own health data.

At HANA, 85% weekly engagement is what we see consistently across deployed clinics. Not because we built magic software. Because outbound voice is structurally the right channel for patients who need follow-up most. Which is most patients. The Rasa platform analysis published this April confirmed what we've been seeing in the field: healthcare voice AI has moved past the experimental stage, and the best systems are now handling real clinical conversations at scale.

Does Voice AI Actually Improve Clinical Outcomes, or Just Engagement Numbers?

Engagement only matters if it connects to something real. Fair question. When you call a patient who discharged three days ago and ask how they're doing, you find things. Medication confusion. Wound concerns. Symptoms they dismissed as normal. Early infection signs they didn't think were worth mentioning. Those findings, when triaged properly, become interventions. And interventions at day three prevent readmissions at day fourteen.

The clinics HANA works with report 31:1 ROI, and a significant portion of that number comes from exactly this: catching the patient who was heading back to the ED in two weeks and rerouting them. Early. The AI call doesn't replace the clinician. It dramatically expands the surface area where the clinician's attention can land.

Is AI-Powered Patient Follow-Up Safe for High-Risk Patients?

Yes, when it's designed with proper escalation protocols. The AI's job isn't to make clinical judgments. Its job is to surface the right information to the right human at the right time. Across more than a million patient interactions, HANA has had zero critical adverse events.

That's not an accident. It's a design choice. Every workflow has escalation paths built in. Every unusual response gets flagged immediately. The system is built to be conservative, not clever, and clever is genuinely the enemy of safe in clinical contexts. The platforms getting this wrong are trying to make the AI do too much. Voice AI that listens, records, summarizes, and escalates creates safety. Voice AI that tries to reassure or advise past its competency boundary creates risk. The distinction is enormous.

How Is AI Voice Different From an Automated Phone Tree?

A phone tree presents options. AI voice has a conversation. That's not a marketing distinction.

When a 74-year-old post-surgical patient calls back, they're not selecting from a menu. They're saying "I've been having trouble sleeping and my leg looks a bit puffy." The system needs to understand that, follow up, ask the right clarifying questions, and flag appropriately. Is "a bit puffy" normal post-op swelling or the beginning of a DVT? The AI doesn't diagnose. But it knows to ask more and escalate what it hears. A phone tree doesn't know to do any of that.

HANA runs on fully open-source infrastructure, self-hosted, with no OpenAI dependency. That's not a technicality. For health systems under compliance pressure, it's often the difference between being able to deploy and not. Your patient conversations stay in your environment. Your data sovereignty stays intact.

Key Takeaways

The mental health app taught me something I needed to learn the hard way. Technology designed to help people doesn't get used just because it's thoughtfully built. It gets used when the friction is lower than the motivation required to engage with it. For most patients, especially the ones who need follow-up most, that motivation is gone. They're tired. They're confused. They're not going to open your portal. But they'll answer the phone.

Eighty-five percent engagement isn't a number we achieved by being clever. It's a number we achieved by being accessible. Show up in people's lives through the channel they already use, in the moment they least expect to need support, and they'll let you in. That's the whole thing.

If your clinic is still measuring patient engagement by app logins and portal opens, you're measuring the patients who don't need the most help. The ones who do are waiting for a call.

FAQ

What is a good patient engagement rate for AI follow-up in healthcare?

A well-implemented AI voice follow-up program should achieve 70-90% weekly engagement. Industry baselines for apps and patient portals typically sit at 15-20%. The gap comes down to channel: outbound voice meets patients where they are rather than requiring them to initiate contact on their own.

How do I know if AI patient follow-up is right for my clinic?

If your clinic has patients who regularly miss follow-up appointments, struggle with medication adherence, or have higher than acceptable readmission rates, outbound AI follow-up is almost certainly worth exploring. HANA's case studies cover clinical contexts from post-surgical follow-up to chronic disease management programs.

Can I try this without a large upfront commitment?

Yes. The best place to start is a direct conversation about your specific clinical context. Book a discovery call here and we'll talk honestly about what realistic outcomes look like and whether HANA is actually the right fit for your situation.