Why Patients Answer AI Calls But Ignore Your Patient Portal
I built a mental health app once, for people living with bipolar disorder. It was a beautiful thing: mood tracking, sleep logs, a chart that showed you your own ninety day pattern in a way that made you go oh, so that's what's been happening to me. We shipped it. We were proud of it.
Fifteen percent of patients opened it in a given week.
Fifteen. And I'm a clinical psychologist, so this wasn't some engineer guessing at what patients want. I'd sat in rooms with these people. I knew their diagnoses, their partners' names, which of them had a dog they'd bring up whenever the session got too hard. They still didn't open the app.
So we stopped asking them to come to us. We called them instead, with AI, and 85% picked up and kept picking up, week after week. That gap between fifteen and eighty five is basically the entire reason HANA exists.
Which is why I read the Mount Sinai paper twice.
What did the Mount Sinai voice AI study actually find?
Researchers at Mount Sinai ran a voice AI assistant called Sofiya through 1,606 pre-procedural calls to 1,431 cardiac catheterization patients, and published the results in npj Digital Medicine in July 2026. Call completion hit 86.4% in the first ninety day phase and held at 87.9% in the second, when nurses ran the thing without developers hovering over it. System errors dropped from 6.0% of calls to 2.6%.
Completed doesn't mean the patient picked up. It means they reached the end of the script and answered every clinical question. Medications, allergies, fasting instructions, the whole prep. Answered.
That's a peer reviewed number, out of a real cath lab, on real patients who skew older and sicker than anyone's app user. Our own outcomes data shows the same shape of curve. Voice does something the screen doesn't.
Why do patients answer AI calls but ignore the patient portal?
Because a ringing phone is an interruption and a portal is an invitation. Interruptions get answered. Invitations get postponed until they're irrelevant, which is exactly what a 2026 secondary analysis found when it looked at patients who'd been enrolled in a text based post discharge program and still ended up back in hospital: fewer than half of them had engaged with the program at all before the return visit.
I learned this properly in Australia, of all places, crossing the desert with a circus (feels like another lifetime, honestly). The performers pulling the biggest crowds weren't doing the hardest acts. Aerial silk is astonishing and technically brutal and people watched it politely, arms folded. Fire chains stopped traffic. Accessible beat complex, every night, in every town.
Patient engagement works the same way. The most sophisticated intervention loses to the one that meets someone where they already are, which for a 71 year old with a stent scheduled Tuesday is a phone that rings.
What does an 87% completion rate do to a clinic's week?
It changes the arithmetic at your front desk. If a coordinator spends four minutes on a prep call and you're running 300 procedures a month, that's twenty hours gone before anybody has picked up an inbound line. Hand that to a voice agent that finishes 87% of them unassisted and you've bought back most of a full time role without hiring anyone.
Our clinics land around 31 to 1 on return, and it's almost never one dramatic saving. It's this, repeated. Prep calls, recalls, post op checks, the small loops that never close because nobody ever had the twenty hours.
Does voice AI only work for procedural prep?
No, and the more interesting thing is buried underneath the Sinai numbers. Structured prep is the easy case. Defined script, known patient, fixed date. The hard cases are open ended: post discharge check ins, chronic care follow up, the patient who needs something they can't quite name yet.
We've run over a million patient interactions across five countries and three languages with zero critical adverse events, and the calls that matter most are hardly ever the scripted ones. They're the one where someone mentions offhand that their leg is swollen and the agent escalates before it becomes an admission. Those show up in the case studies far more than the tidy ones do.
What should a clinic check before letting voice AI near patients?
Three things, and none of them are the demo.
Where does the call go when a patient says something frightening? If the answer is a queue, walk away. Escalation has to be immediate and it has to have a name attached. Second, where does the audio live? Ours is open source and self hostable precisely because plenty of clinics can't ship patient voice data to somebody else's API and shouldn't have to. Third, who writes the clinical logic? If it's the vendor's prompt engineer and not your nurse, you'll find out in month three.
Key takeaways
The Mount Sinai result isn't remarkable because AI worked. It's remarkable because it held up after the developers stepped back and the nurses took over, which is the phase most pilots never survive. Voice beats portals for the same reason a knock beats a letter. And the clinics getting real value out of this aren't the ones running the fanciest model, they're the ones who picked the boring, repetitive, twenty hours a month loop and closed it properly. Start there. The exciting stuff gets easier once the dull stuff is handled.
I still think about that fifteen percent. They weren't disengaged patients. They were patients we'd asked to do the wrong thing.
Frequently asked questions
How accurate is AI voice for pre-procedural patient calls? In the Mount Sinai study, system errors occurred in 2.6% of calls during the nurse led phase, down from 6.0% during the initial stabilization period. Accuracy improves when the agent is tuned against your own patient population rather than a vendor benchmark.
Do older patients accept AI phone calls? Generally yes, and more readily than they accept digital tools. The cath lab population skews older and still completed close to 88% of calls, while portal and app engagement tends to fall off sharply with age.
Can voice AI replace nurse follow-up entirely? No, and it shouldn't. It absorbs the repetitive structured portion so nurses spend their hours on clinical judgment and escalations. The Mount Sinai authors described it as augmenting routine tasks, which is the honest framing.
If your team is still doing prep calls by hand and bleeding hours into it, grab a slot and we'll walk through your numbers together.