Retell’s voice agents hit 100 million calls a month, by a vendor’s count

Contact centers are where AI meets the public at volume, and Retell says it now handles more than 100 million calls a month for over 4,000 businesses, up from about 30 million early this year. When one platform more than triples call volume in under a year, voice agents stop being a pilot and start being infrastructure. Retell’s harness manages audio, turn-taking and interruptions, while OpenAI models generate responses and request tool calls that Retell executes against customer systems. It uses regional API endpoints for latency and data residency and runs a backup telephony stack for redundancy. Other OpenAI-powered voice platforms report similar traction, such as Ringg resolving up to 65% of calls (Ringg).

Retell picks models on its own benchmark built from hard, multi-turn production scenarios and reruns it when new models arrive. Choosing on real calls rather than demos is the discipline this market has lacked. Customer outcomes are concrete: Anker reports an 80.4% case resolution rate on Tier-1 post-sales support with an NPS of 63, and Pine Park Health says it recovered medical-assistant time equal to 2.1 full-time roles and moved it to patient care.

The frame matters, though. This is a customer story published by the model vendor, and the benchmark is Retell’s own. GPT-6 Astra ranking first at 89.7% for instruction following is a useful data point, not an independent verdict; the same comparison from someone selling neither side would carry more weight.

The handoff design is the best part. When a human is needed, the agent briefs the representative, answers their questions and confirms they are ready before connecting the caller, so the machine gathers context and the person exercises judgment. The worry is the other case, where the roughly 20% of calls that do not resolve get trapped in a loop with no exit. Retell’s engineers also use computer-use models to click through and try to break front-end pull requests, a testing pattern worth copying.

The revenue-per-employee figure cuts both ways. Retell’s engineering team leans on Codex for research, building and debugging, and the company runs at $1.6 million in annual revenue per employee. Retell says it is still hiring, just not at the pace of the business. If that spreads, the effect is less overnight layoffs than a quiet flattening of entry-level engineering and support roles, jobs that never get posted. Coding agents are stretching into other functions too (Amp beyond code). More in our OpenAI coverage, agents coverage and business coverage.