Retell AI Built a $60M Business on 5 Cents a Minute

The phone was supposed to be dead. Every product deck for a decade said so - chat first, then bots, then a contact form buried three clicks deep. Meanwhile the actual phone lines kept ringing, kept going to voicemail, and kept getting handed to call centers half a world away. Retell AI looked at that gap and did something unfashionable. It answered the phone.

The company sells AI voice agents - software that picks up a call, talks like a person, books the appointment, checks the account, and hands off to a human when it needs to. It reached roughly $60 million in annual recurring revenue by April 2026, up about 650% in a year, according to research firm Sacra. The part that makes other founders sit up: it did that on about $5 million in total funding. In a market where AI infrastructure companies raise nine figures before they ship, Retell got to a $60M run rate on a seed round.

01 / The MeterThe pricing is the product

The most interesting thing about Retell AI is not the voice. It is the meter. The platform charges roughly $0.055 per minute of conversation, plus pass-through costs for the underlying speech-to-text, text-to-speech, and language models. No per-seat license. No annual enterprise contract that takes six months to sign. You pay for calls the way you pay for electricity - by how much you actually use.

Put that next to the alternative and the whole business snaps into focus. An offshore human contact-center agent costs somewhere between $0.30 and $0.80 a minute once you count wages, management, and overhead. Retell's customers report saving 70 to 95% per minute. That is not a discount. It is a re-pricing of an entire labor category, one call at a time.

When the unit economics are that lopsided, the sales pitch changes. You are not asking a company to believe in AI. You are asking them to do arithmetic. Everyone who runs a phone-heavy operation - a lender chasing payments, a clinic confirming appointments, an insurer taking first-notice-of-loss calls - already knows exactly what a minute costs them. Retell just moved the decimal point.

02 / OriginsFrom dubbing to dial tone

Retell AI did not start as a phone company. Bing Wu, Todd Li, Zexia Zhang, Weijia Yu and Evie Wang first built a tool for live-streamed dubbing - translating creators for international audiences in real time. The through-line to what they do now is the same hard problem: make a machine talk like a person, fast enough that you forget it is a machine. When they saw how badly call-center-heavy businesses wanted that, they pivoted.

The founders came from ByteDance, Meta and Google - companies that learned to serve products to billions of people. Wu, the CEO, spent three years at ByteDance leading products at consumer scale. They pointed that instinct at the least glamorous surface in all of software: the phone call. Y Combinator rejected them the first time. They sharpened the demo and got into the Winter 2024 batch, because a working voice agent sells itself in about 30 seconds - you call it, it answers, and the pause where you expect awkwardness never comes.

By the time the platform drew press attention in 2024, it already had hundreds of paying customers and named names like the telehealth company Ro. OpenAI featured it as a builder on GPT-4o. The agents answered in roughly 600 to 800 milliseconds and could handle interruptions - the small, human thing where you cut in mid-sentence and the other side adjusts.

03 / The LayerSell the control, not the model

Here is the strategic bet that keeps Retell relevant while the models underneath it change every few months: it does not sell you the model. It sells you the control layer - the thing that orchestrates the call, calls your functions, checks your knowledge base, and decides when to transfer to a human. Underneath, developers can swap speech and language providers at will: Deepgram, Cartesia, ElevenLabs, OpenAI, Anthropic. If a cheaper or better voice ships next quarter, you switch a setting, not a vendor.

In an AI market that reprices monthly, that optionality is the product. Retell is deliberately model-agnostic and developer-first, which is a different shape from the managed enterprise platforms. It hands you an API and gets out of the way. You can wire up a production voice agent in an afternoon.

04 / The GiantsMicrosoft, Salesforce and Cresta want the same calls

This is where it gets interesting, because Retell is not alone on the line. Microsoft is pushing voice AI through Dynamics. Salesforce is selling Agentforce into its enormous installed base. Cresta blends real-time human-agent coaching with autonomous AI, built on its own model, aimed squarely at big enterprise contact centers. Each of them has something Retell does not: distribution, a suite, a logo the CIO already trusts.

What Retell has is a smaller surface area and a sharper edge. The incumbents bundle voice into a broader platform and sell it by the seat to the enterprise. Retell sells infrastructure by the minute to the developer. Those are different customers with different clocks. The enterprise buyer wants one throat to choke; the developer wants to ship this week and pay for what runs.

There is also a nearer cohort - Vapi, Sierra, Decagon, Bland AI - fighting over the same developer-to-mid-market ground. Sierra and Decagon lean toward managed services for large accounts. Retell's counter is to stay lean and usage-priced, and let the meter do the selling. So far the customers agree: BPOs like Everise, lenders like Sunshine Loans, and insurers like Matic run real volume through it, in regulated verticals where a dropped call has consequences.

05 / What You Can Do With ItThe practical part

Strip away the funding numbers and here is what a business actually gets. You point Retell at a phone number and describe the job in prompts and function calls. The agent can book appointments, navigate an IVR, look up answers from your knowledge base, run outbound batch campaigns from a branded caller ID, and transfer to a human when the conversation needs one. After the call, it hands you transcripts, post-call analysis, and automated QA on every conversation - not the sampled 2% a human QA team gets to.

The verticals that bite hardest are the ones drowning in repetitive calls: healthcare scheduling, financial services, insurance, logistics, retail, travel, and debt collection. One healthcare client reported lifting scheduling NPS by 38%. An EV support operation cut support costs by more than half. A collections firm runs roughly $280,000 a month in recovered payments through agents without torching customer trust. None of that requires believing the phone came back to life. It just requires answering it every time.

The honest caveat: voice AI is still young, and "human-standard" is a claim, not a guarantee. Accents, cross-talk, and genuinely upset callers remain hard. Retell's own framing - "built to scale, human by design" - is a design goal as much as a description. What is not in dispute is the trajectory: $60 million in run-rate revenue, growing 650% a year, is the market voting with its phone lines.

The lesson other founders keep pulling out of this one is not really about voice at all. It is that you can build a large, fast-growing software business without a mega-round if you charge for something people already want more of, price it against a cost they can already see, and refuse to lock yourself to a model that will be obsolete by spring. Retell AI did the unfashionable thing and answered the phone. The revenue suggests plenty of people are glad someone finally picked up.