Six numbers from running a voice AI in production

Canarics runs realtime AI voice agents on live business lines across the Baltics. Most writing about voice AI comes from demos. These numbers come from production traffic, and several of them were unwelcome. Sample sizes included; the aggregate data lives on our data page.

49.5%of engaged callers get resolved with no human at all

n = 1,060 engaged conversations (caller spoke at least twice), 90 days

Of the callers who do engage, half finish their whole errand with the AI: bookings, order status, routine questions. Another 8.7% get transferred to an employee, live, with context. The rest either got an answer that needed no logged action or gave up somewhere in the middle. We reported 48.5% for a single client earlier; measuring across the fleet independently reproduced it, which is the kind of boring confirmation you want from a metric.

72sis the median engaged conversation, almost exactly a human call

n = 1,060 AI conversations vs 19,893 human calls on the same lines

Median engaged AI conversation: 72 seconds. Median human call on the same lines: 59 seconds. The picture of AI calls as either ten-second gimmicks or ten-minute hostage situations is wrong on both ends. When the AI can actually do the task, the conversation takes about as long as it would with a person.

6–8%of short utterances are lost, and better audio does not fix it

measured across controlled test call series, 2026

We spent weeks on a theory that lost caller utterances ("yes", a name, a number said quickly) were a transport problem, and rebuilt the audio path to deliver lossless wideband audio end to end. Delivery improved measurably. Capture did not: 6 to 8% of short utterances were still missed. That is the current floor of realtime speech models, not of telephony plumbing.

The engineering consequence: stop optimizing the pipe and design the conversation to survive losses. The agent must ask again naturally instead of guessing, and anything written to a database must come from the caller's confirmed words, never from a guess about a half-heard one.

~9%of "answering machine" detections are actually humans

false-positive rate 8.7%, measured on our outbound traffic

Premium answering-machine detection mislabels a live human as a machine roughly once in eleven detections: slow speakers, noisy workshops, unusual greetings. If your agent hangs up whenever the carrier says "machine", it hangs up on real people at scale, and outbound calling at scale means those people remember.

Our rule: machine detection may never terminate a call by itself. The agent verifies conversationally, and once a human has spoken even a single time, no automated signal is allowed to end that call. We measured the false-positive rate and then treated the detector as a hint, not a verdict.

250msper turn was worth chasing, and the next 250 is not

We cut roughly a quarter second from every conversational turn by tuning silence detection and audio padding. Below a certain threshold the tradeoff inverts: trim more aggressively and the agent starts talking over slow speakers and clipping the first syllables of fast ones. Latency work in voice AI has a floor set by human pause behavior, not by compute, and the complaints tell you when you have crossed it.

6%of callers switch between Russian and a Latin-script language mid-call

n = 798 analyzed conversations, caller turns only, 90 days

One caller in sixteen crosses the Cyrillic-Latin language boundary inside a single conversation, and that is only the switches a script-level classifier can prove; Latvian-Lithuanian-English drift is invisible to it and comes on top. In the Baltics, mid-call language switching is not an edge case, it is Tuesday. An agent that needs a language selected up front fails these callers on the turn where it matters.

What we would tell ourselves a year ago