The automation number gets the attention. The routing number earns the trust. And the 2% the system gets wrong? That turned out to be the most valuable number of all.

The call came in a few minutes after nine on a weeknight. A patient at a regional primary care group, describing symptoms that couldn't wait until morning. A year ago, that call would have gone to an after-hours answering service that could take a message but couldn't do much else — or, during business hours, into a hold queue where waits stretched toward half an hour and roughly one in ten callers simply gave up before reaching a person.

Instead, the system recognized within seconds that this wasn't a scheduling call. It flagged the acute language, skipped the booking flow entirely, and routed the patient to the on-call clinical team. Total wait: about as long as it takes to read this sentence.

When people ask me about this deployment, they ask about a different number. They ask about the 40%.

At a glance: 98% of inbound calls handled by the system · 40% resolved fully autonomously · 58% intelligently routed to the right human · 98% accuracy on core tasks — and a 2% rate that feeds the roadmap

The number everyone asks about

Forty percent of this group's inbound calls are now handled start to finish without a human. Scheduling, rescheduling, confirmations, routine questions — the patient calls, the AI answers, the appointment lands in the EHR, and nobody's queue gets longer. At this group's call volume, that's thousands of conversations a month that never touch a hold line. It runs around the clock, in English and Spanish, across every location.

It's a genuinely good number. It's the one that makes executives lean forward in a QBR. And after eighteen months of deploying these systems, I've come to believe it's the second most important number on the dashboard.

The number that should impress you more

The 58% is the one I watch. Those are the calls the system handled by knowing, within seconds, that it shouldn't finish them — and getting them to the right human on the first try.

That sounds like the unglamorous half of the work. It's actually the hardest. A patient who mentions chest pain or low blood sugar needs a nurse, not a scheduling flow, and the system has to hear the difference in the opening seconds of a call. A Spanish-speaking patient needs first-line support in their own language, not a transfer and a callback. A caller who dialed the primary care line but needs the specialty clinic across town needs to land there on the first hop, not after two hold queues and a sigh.

Anyone can automate the easy calls. The discipline is in knowing — in seconds — which calls you shouldn't.

Routing is where the clinical stakes live, so routing is where the trust lives. The ops leads I work with didn't start trusting the system when it booked its thousandth appointment. They started trusting it when they listened to the acute calls and heard it get out of the way, fast, every single time.

What the 2% teaches you

The system runs at 98% accuracy on its core tasks. The honest question — the one good operators always ask — is about the other 2%.

Here's what we do with it. Every miss gets isolated and categorized: a name the speech model fumbled, a scheduling rule with an edge case nobody documented, an insurance nuance specific to one location. Then we sit down with the client's operations team — the people who live in the call volume every day — and walk through them together. What should have happened? What rule was missing? What does the fix look like?

A 2% failure rate without a feedback loop is just failure, repeating quietly forever. A 2% failure rate with a feedback loop is the most precise product roadmap you will ever get. Several of the system's best routing behaviors today started life as a miss that an ops lead caught, flagged, and helped us fix.

98% accuracy isn't a finish line. It's a feedback loop.

From replacing humans to freeing them

Here's the part of this story nobody puts on a dashboard. The front desk team at this group didn't shrink. Their work changed shape.

Before, their day was triage by exhaustion: a ringing queue, a half-hour backlog, patients who'd been on hold so long the conversation started in frustration. Now the calls that reach them are the ones that genuinely need them — the complex cases, the distressed patients, the judgment calls, the insurance puzzles that require a human who knows the practice. The easy calls are gone, and what's left is the work they were actually hired to do.

That, more than any percentage, is what "AI handling your calls" should mean. Not an empty desk. A desk where the phone ringing means something.

The better question

The question I get most often is some version of "how much can the AI handle on its own?" It's a fine question. But after watching this deployment run, I'd tell any operator evaluating these systems to ask a different one: what happens to everything it can't?

Where do those calls go, and how fast? Who reviews the misses, and with whom? Does the 2% feed a roadmap, or a shrug?

The automation rate tells you what the system can do today. The answers to those questions tell you what it will be doing a year from now. The goal was never fewer humans on the phones. It was the right human, on the right call, faster — and a system that gets better every time it isn't.

Tom Nork is Head of Client Experience at Parakeet Health, where he leads implementation and client partnerships across large specialty and primary care groups.