Call volume doesn’t show whether patients get on therapy. These three benchmarks, which Infinitus is sharing publicly, measure what does: how AI holds up in real conversations, with payors, and in clinical judgment.
The care exists. But between a patient and that care sit layers of friction. Even when AI is implemented in critical workflows, small failures like a misheard word or background noise can decide whether patient care actually moves forward.
AI can help remove that friction, but only if it performs reliably in the situations that matter. It’s not enough to show what the technology could do in a polished demo. We need to know what it actually accomplishes in the real world.
That’s why we’ve introduced three new benchmarks, announced at our BEYOND 2026 event in New York on Nov. 12, 2026. These are standards we’re holding ourselves to, because we believe healthcare AI needs a higher standard of proof. We’ll share our results openly and on an ongoing basis.
Announcing three benchmarks measuring safety and success
Healing Rate measures how often an AI agent recovers when a conversation goes off course, whether from interruptions, background noise, or misheard words, and still completes the task. Our agents get back on track and finish the job 95% of the time. Alongside Healing Rate, we measure whether agents say what patients need to hear correctly. Medication names and dosing information are each pronounced accurately 99% of the time.
Agent Navigation Rate tracks how reliably our agents work through the phone systems of major US payors and reach the right place. They succeed 92% of the time. When we compare against people calling the same payor lines, humans reach the right person significantly less often, and wait on hold longer. And remember, every minute stuck in a phone tree is a minute added to a patient’s wait.
Clinical Severity Classification looks at how well AI recognizes clinical need, and when it brings in a person. We measure our agents by evaluating how well they perform the same tasks as licensed nurses. Our system matches the consensus at least as often, and in some types of cases more often, than any individual clinician.
Why we’re publishing these numbers
Healthcare has seen plenty of AI claims and very little evidence. Buyers are asked to trust systems with workflows where a mistake can delay patient care, with little more than a demo and a promise. We think that gap is the real barrier to adoption, and the way to close it is to publish what we measure and how we measure it.
That’s why we’re sharing these benchmarks. No AI system is right every time. So we measure recovery in real calls, not just first-pass accuracy, and we’re publishing the results so you can judge them yourself.
Questions to ask any AI vendor
- What happens when a caller interrupts, or the line is noisy?
- How reliably does your agent get through payor phone systems, and how long does it wait?
- How accurately does it pronounce drug names and doses?
- What standard do you hold it to when it decides a patient needs clinical help?
- Can I see the results, and how often are they updated?
Want to go deeper?
Check out the deep dives on Healing Rate, Agent Navigation Rate, and Clinical Severity Classification, or talk to our team to learn more.
