Why Do Patients Ignore Your Portal and Answer an AI Phone Call?

There was a day, years ago, when I watched a Stripe dashboard cross a million dollars. One day. One product. I remember refreshing it like an idiot, and I remember the specific feeling, which wasn't joy exactly, more like vertigo.

The product wasn't good.

I know that now and honestly I half knew it then. Scale doesn't fix a bad product. It just makes the bad product louder, and hands you a bigger number to hide behind while churn quietly eats the thing from underneath. (took me about eighteen months and an unreasonable amount of money to learn that one)

I think about that every single time a clinic tells me their patient portal has ten thousand registered users.

Why don't patients use the patient portal?

Because a portal asks the patient to remember you exist. That's the entire mechanic. It sits there, passive, waiting, and it works best for the patient who was already organised enough not to need it.

The research keeps confirming this in ways that should be embarrassing. The MORE-PC trial, one of the largest post-discharge SMS studies ever run, found no significant reduction in 30-day readmissions from automated text messaging. The patients who did engage skewed younger and commercially insured.

So the tool performed best for the people least likely to bounce back.

What's the difference between reaching a patient and offering to be reached?

One is a push. The other is a hope.

We built the push version after the hope version failed. Our first product was an app for people with bipolar disorder. Gorgeous, clinically sound, 15% engagement. Then we picked up the phone instead, with a voice agent doing the calling, and engagement went to 85%. Same patients. Same clinical content. Different direction of travel.

A heart failure outreach programme published this year split its cohort by engagement and found 7.7% readmissions among the highly engaged versus 21.6% among everyone else. Same programme. The variable was contact, not content.

That's not a technology insight. It's a behaviour insight that happens to need technology to run at volume.

Which follow-up calls are actually worth automating?

Start with the ones your staff already skip. Not because anyone is lazy. Because there are four hundred of them and four people.

Post-op day one and day three checks. Medication adherence at week two. No-show recovery inside the hour, while the slot is still fillable. Pre-visit intake so the appointment starts at minute zero instead of minute nine. Chronic care management touchpoints that are billable and currently unbilled. We break these down by specialty in our use case library, and the pattern never changes: the highest-value call is the one nobody has time to make.

Your no-show rate isn't a scheduling problem. It's a staffing arithmetic problem wearing a scheduling costume.

Does an AI voice call feel cold to patients?

Less than you'd think, and considerably less cold than a portal notification. A call arrives, a voice asks how the knee is doing, the patient says it's swollen and warm, and the system escalates to a nurse.

Total elapsed time: ninety seconds. Nobody logged in to anything.

Across more than a million patient interactions in five countries and three languages we've had zero critical adverse events, and the thing patients comment on most isn't the AI. It's that somebody checked. Others are documenting the same shift in post-discharge workflows, and it keeps landing in the same place. The emotional unit of care isn't the sophistication of the tool. It's whether the phone rang.

What do you need in place before this works?

Less than most vendors imply. An EHR you can read from and write to. A clinical escalation path with an actual human at the end of it. And a decision about which conversations you're comfortable delegating. That's the honest list, and the integration docs are public if you want to see exactly what it touches.

The part everyone skips is the escalation path. If your AI flags a red symptom at 7pm and it lands in a queue nobody opens until Tuesday, you haven't built a care programme. You've built a liability with good branding. Decide who catches the ball before you throw it.

What does this do to the numbers?

Clinics running voice follow-up with us see roughly 31:1 return, a ratio I'd have rolled my eyes at before I ran the arithmetic myself. It isn't magic. It's recovered no-shows, billable care management that was being left on the table, and readmissions that never happened because somebody caught a wound infection on day three instead of day nine. We publish pricing openly rather than hiding it behind a demo, because the maths should be checkable by you, not performed at you.

My daughter is ten. I told her once that I work for her, not the other way around, and she accepted it as completely obvious, as she should. Same logic here. The technology works for the clinic. If the clinic is working for the technology, something is inverted, and no dashboard is going to unbend it.

Key Takeaways

Patient portals and SMS programmes underperform for a structural reason, not a design reason: they require the patient to initiate, which quietly filters out the exact population most at risk. Outbound voice inverts that. The engagement gap between passive tools and active calling isn't marginal, it's roughly 15 to 20 percent versus 85 percent, and engagement drives every downstream outcome you care about. Start with the calls your team already can't make, build a real escalation path before you automate anything clinical, and measure contact rate instead of registered users.

If your portal has ten thousand users and your follow-up completion sits under a third, you don't have an engagement programme. You have a mailing list.

FAQ

Will patients hang up on an AI phone call? Some will, the same way some hang up on a human scheduler. In practice completion rates run far above portal or SMS engagement, because answering a phone requires no account, no password, and no memory of a login. The call comes to them.

Is voice AI for patient follow-up HIPAA compliant? It can be, and it depends entirely on architecture. HANA is open-source and self-hostable with no OpenAI dependency, which means patient data can stay inside your own infrastructure rather than crossing into a third-party model provider. Setup is documented publicly.

How long does deployment take? Weeks, not quarters, for a single use case like post-op follow-up or no-show recovery. The long pole is almost never the technology. It's agreeing internally on the escalation path and who owns the flagged calls. Real timelines are in the case studies.

Want to work out which single uncalled patient is costing your clinic the most money right now? Grab twenty minutes.