Real calls are messy. Our agents recover from interruptions, noise, and errors to complete the call 95% of the time, with 99% accuracy on medication and dosage pronunciation.
Picture a patient calling an information line about a new prescription that’s helmed by an AI agent. She’s anxious, and every time the agent starts an answer, she talks over it. A truck rumbles past and the line crackles. She reads out a member ID, and one digit comes through wrong.
In the first call, the agent loses the thread. It can’t finish its questions, so it can’t finish the task. The patient hangs up and calls back, and has to explain everything again.
In the second version, the agent holds its place. It acknowledges her worry, politely asks the question again, catches the wrong digit, and confirms the right one. The call ends with everything completed and without the patient needing to call back.
The difference between those two calls is what Healing Rate measures. We all want to be the second call, and 95% of the time, we are.
The Infinitus Healing Rate
Every voice AI demo sounds great. The caller speaks clearly, answers every question, and nobody ever interrupts the agent.
But real calls aren’t like that. If a patient is anxious and keeps talking over the agent, the questions that move the call forward might never get asked. A caller might mumble an NPI or BIN. A dog barks in the background. The speech-to-text hears “B” instead of “D.” Each is a point where an AI agent can lose the thread.
Healing Rate measures how often an agent recovers from these external errors, brings the conversation back on track, and still completes the call. In our measurements, our agents recover and finish the job 95% of the time.
What Healing Rate examines
Interruptions: A patient calls about a new prescription, and she’s anxious about the cost. Every time the agent starts to ask for her date of birth, she cuts in with, “But how much is this going to be?” The agent can’t skip the question, because nothing else can happen until it knows who it’s talking to. So it acknowledges the worry, promises to come back to it (“I’ll get to the cost in just a moment, I just want to make sure I have the right account first”) and asks again. A recovery here means the call finishes with her identity confirmed and her cost question answered, in that order.
Background noise: A patient calls from a pharmacy counter, with announcements overhead and a line behind her. The agent catches half of a medication name and nothing else. It doesn’t guess, and it doesn’t restart the whole call. It repeats back what it did hear (“I caught the first part of that, could you say the name once more?”), waits out the noise, and picks up where it left off. If the line stays too loud, the agent can offer to call back at a better time instead of pushing through with bad information.
Speech-to-text errors: A provider’s office reads out a member ID “B as in bot, four, zero, one.”
The transcript hears “D as in dot,” or as another example, “fifteen” comes through as “fifty.” These are small mistakes, but one wrong character can derail an entire benefit verification. The agent reads the number back in short groups, the caller corrects it, and the call continues with the right data. What counts is that the error gets caught before it reaches the record, not after.
In all three cases, the agent doesn’t need a perfect call to finish the job. It just needs to notice that something went wrong and get back to the information it needs.
Getting the words right
The common thread is outcomes. Recovering isn’t only about staying calm. It’s about getting the information needed to move the conversation forward and finish the job. But recovering from errors is only half the story. What the agent actually says also has to be right, so we benchmark medical accuracy alongside Healing Rate. Our agents pronounce medication names accurately 99% of the time, and dosing and related information just as often.
We measure this against a dictionary of 10,000-15,000 medical terms. We sample from it, generate the agent’s speech, and have humans annotate it with help from our clinical teams. The goal is a text-to-speech vendor and configuration that pronounces medications the way a doctor would and a patient would understand. Dosing gets the same human-reviewed standard because misreading “one tabled twice daily” would be a critical safety issue.
As part of related information, we also measure identifier accuracy, since NPI numbers, zip codes, and phone numbers are where AI has struggled historically, and even one wrong character can completely derail a benefit verification.
Why it matters
When an agent can’t recover, the work falls to the patient. They have to call back, repeat themselves, or just wait through another delay in getting their treatment. At 95%, that burden stays with the agent instead of landing on the person it’s supposed to help.
For your team, that means fewer failed calls, less rework, and fewer patients stuck in loops.
Read up on more Infinitus benchmarks, or talk to our team about how our agents can handle your calls.
