Voice AI Vendors Promise 50% Booking Lifts. Where's the Evidence?

A clinic owner forwarded me a vendor deck last week. Big claims. 40% more bookings, 50% fewer no-shows, a hockey stick on every slide. She asked me one question: is any of this real?

Honestly, fair question. Deepgram just published a piece asking exactly this: vendors promise 30 to 50% booking lifts, and there are basically zero peer-reviewed studies behind those specific numbers. I run a voice AI company. You'd expect me to be annoyed by that article. I'm not. I think it's the most useful thing published in this space all year.

Why is there so little published evidence for voice AI in clinics?

Because the incentives point the other way. Running a proper study takes 12 to 18 months, and the market is moving in 6-week cycles. So vendors ship case studies instead, and case studies are marketing with a methods section missing. I'm guilty too. We publish our numbers, 85% weekly patient engagement against a 15 to 20% industry baseline, and you have every right to ask how we measured them. We try to answer that openly in our research notes. Most decks won't.

I learned this lesson the expensive way. Before HANA I built a mental health app for bipolar patients. Beautiful dashboards. 15% engagement. The data I showed investors looked great because I chose what to show. The patients told the truth by ignoring the app.

What do the independent studies actually show?

The adjacent evidence is stronger than the vendor evidence. A 2026 quasi-randomized trial of nurse-led post-discharge calls found a 28% relative reduction in 7-day ED visits. Calls work. Humans just can't make enough of them. That's the honest pitch for voice AI: not magic, just coverage. The same call, made every time, to every patient.

Meanwhile a 2026 industry field guide reports no-show reductions of 10 to 20% from reminder calls. Plausible. Bounded. Believable precisely because it isn't 50%.

What should a clinic owner ask a voice AI vendor?

Ask for the denominator. Always the denominator. "85% engagement" means nothing until you know: of how many patients, over what period, counting what as engaged. Then ask three more things. Can I talk to a clinic that churned? What happens to my data if I leave? What did your worst deployment look like?

In the circus I toured with across the Australian desert (yes, really, another lifetime), the acts that drew the biggest crowds were the simplest ones. Fire chains, not aerial silk. Vendors selling complexity are usually hiding something. The good ones sell something boring and provable.

How do you run your own evidence instead of trusting theirs?

Pilot with a control group, even a scrappy one. Pick one workflow, say recall calls for lapsed patients, and split your list. Half get the AI calls, half get whatever you do today. Run it for 60 days. Count bookings, not conversations. We structure pilots exactly like this and walk through the design in our case studies, because a clinic that measures honestly stays longer. The ROI we see, around 31:1, only matters because the clinic computed it from their own numbers, not ours.

Sixty days. One workflow. Your data.

Does the missing evidence mean you should wait?

No. It means you should buy differently. The staffing crisis at clinic front desks isn't waiting for peer review, and the patients who never pick up your one overworked receptionist's call aren't either. The technology is production-ready for bounded workflows: reminders, intake, recall, follow-up. See the use cases that actually ship. What's not ready is buying on a slide deck's promise.

Key takeaways: the published evidence for specific vendor claims is thin, and the vendors know it. The adjacent clinical evidence says structured outbound calls genuinely reduce ED visits and no-shows. The fix isn't waiting, it's piloting with your own control group, demanding denominators, and treating any unverifiable claim as decoration. Skepticism isn't the enemy of adoption. It's how adoption survives contact with reality.

FAQ

Are there any peer-reviewed studies on voice AI booking lifts?

Not on the specific 30-50% vendor claims, as of mid-2026. The strongest related evidence comes from studies of structured human follow-up calls, which show meaningful reductions in ED visits and no-shows. Treat vendor percentages as hypotheses to test, not facts.

What's a realistic engagement number for AI patient calls?

Depends entirely on the workflow and population. We see 85% weekly engagement in our deployments, against a 15-20% industry baseline for app-based tools, and we publish how we measure it. Anyone quoting a number without a denominator is selling, not reporting.

How long should a voice AI pilot run before deciding?

Sixty days is usually enough for one bounded workflow with a control group. Long enough to see a real pattern, short enough that you're not married to the vendor. If you want help designing one, book a discovery call and I'll walk you through our pilot template.