Vapi's approach to voice agents

A voice agent is an AI agent that works over a phone call or a browser microphone instead of a chat window. It listens, decides what to say or do, and answers aloud, often while calling tools such as a booking system. If the general idea is new, start with our explainer on what AI agents are and how they use tools.

Vapi, a San Francisco company, does not sell a finished receptionist or sales bot. It sells the layer underneath, for developers who want to build their own. This profile relies on Vapi's site and docs, its funding announcements, one news report and an FCC ruling. We have not built or tested an agent on Vapi.

What Vapi makes

Vapi's documentation calls it "the developer platform for building voice AI agents" and says: "We handle the complex infrastructure so you can focus on creating great voice experiences."[1] Developers build with two primitives. Assistants use one system prompt plus tools, for jobs such as support, lead qualification and booking. Squads chain several specialised assistants and pass context between them when the call moves on.[1]

The company was founded by Jordan Dearsley and Nikhil Gupta, who had gone through Y Combinator with an earlier productivity startup, TechCrunch reports.[2] Vapi says it grew out of an AI therapist Dearsley built for himself in mid-2023 and launched publicly on Product Hunt in March 2024.[3]

Vapi's own figures, which we have not verified: more than 1 billion calls handled and more than 1 million developers on the platform.[4][3] It names Amazon Ring, Intuit, ServiceTitan and New York Life among enterprise customers.[4]

Funding. Vapi announced a $20 million Series A led by Bessemer Venture Partners in December 2024.[5] In May 2026 it announced a $50 million Series B led by Peak XV, with M12, Kleiner Perkins and Bessemer participating, bringing total funding to $72 million.[3] The company attributes the round to "a 10x in enterprise ARR growth". TechCrunch, citing a person familiar with the matter, put the post-money valuation at around $500 million.[3][2]

How a voice agent pipeline works

"Every Vapi assistant combines three core technologies," the docs say: speech-to-text, a large language model and text-to-speech.[1] The developer chooses a provider for each from a long list that includes OpenAI, Anthropic, Google, Deepgram and ElevenLabs.[1] Telephony carries the audio to and from the caller.

Vapi's latency guide breaks one conversational turn into five stages that run in sequence:[6]

1. Transport: the caller's audio travels over the network.
2. Endpointing: the system decides the caller has finished speaking.
3. Transcription: speech becomes text.
4. Model response: the model starts replying, measured as time to first token.
5. Speech synthesis: the voice provider starts producing audio, measured as time to first audio.

The guide is unusually candid about its own numbers. Vapi's dashboard total adds the median latency of the transcriber, model and voice only. It "excludes endpointing and transport time", although "endpointing can account for a meaningful share of that delay".[6] The homepage advertises "<500ms average latency" and the docs mention "Sub-600ms response times";[4][1] the same guide warns that latency figures should be compared "only when their scope and conditions match".[6]

AI Chronicle

Turn-taking is where much of the "feel" lives. A start-speaking plan sets how long the agent waits after the caller stops (0.4 seconds by default) and can use a smart-endpointing model to judge whether a pause is mid-thought. A stop-speaking plan sets how much caller speech makes the agent stop talking and how long it waits before resuming.[7] The trade-off is snappiness against interrupting people.

Vapi also supports OpenAI's speech-to-speech models, which process audio in and audio out natively rather than chaining three separate services.[1]

The developer model and pricing

Vapi is API-first, with a web dashboard, SDKs, a command-line tool and "config as code" kept in git.[1] Teams that already pay a model or voice provider can add their own keys, in which case that provider bills them directly.[1] For phone lines, Vapi offers one free US number that handles inbound calls only; outbound calling needs a number imported from Twilio, Vonage or Telnyx, and SIP trunks are supported.[1][8]

Published pricing, as of 16 October 2026, has two layers.[9] Every account pays usage: a Vapi hosting fee of $0.05 per minute, plus transcriber, model and voice costs, which Vapi says it passes through "at cost". Telephony transport is charged by the carrier; Vapi lists its own telephony and web transports as free. On top, optional Success Packages add support and limits:

Add-ons include HIPAA-eligible handling at $2,000 per month and extra concurrent lines at $10 each per month.[9] Existing customers on legacy pricing keep their old terms, the docs note.[10]

What to weigh before deploying

Outbound calls and US robocall law. In a declaratory ruling adopted on 2 February 2024, the FCC confirmed that the Telephone Consumer Protection Act's limits on "artificial or prerecorded voice" calls cover AI-generated voices.[11]

“Callers must obtain prior express consent from the called party before making a call that utilizes artificial or prerecorded voice simulated or generated through AI technology.”

— Federal Communications Commission, Declaratory Ruling FCC 24-17, para. 5.[11]

Telemarketing calls need prior express written consent. Every AI-voice message must identify the business responsible, and telemarketing messages must offer an opt-out.[11] The FCC also said the TCPA allows no "carve out of technologies that purport to provide the equivalent of a live agent".[11] Vapi's own guide sets out the same consent tiers and states the TCPA does not apply to inbound calls.[12] This is US law; other countries have their own rules.

Recording consent. Vapi's docs say "call recording laws vary by jurisdiction".[13] Its automated recording consent plan, which asks permission before recording starts, is an Enterprise feature; other customers build their own consent step.[13]

Disclosure. The FCC rule requires callers to say who is calling. Telling callers they are talking to an AI is a separate question that our sources do not settle. Our view: say it at the start.

Escalation. Vapi's transfer tool can route a live call to a phone number or SIP address, either blind or "warm" with a message or summary for the person picking up.[14] Its own example prompt says to "only transfer after the caller agrees".[14] Who answers when the agent gets it wrong is a separate question, covered in our analysis of who is accountable when an AI agent acts.

A worked example: an after-hours clinic line

Hypothetical. No real clinic, configuration or result.

A dental clinic wants an agent to answer calls after 6 pm, book routine appointments and pass urgent calls to the on-call dentist. Because calls are inbound, Vapi's guide says the TCPA does not apply.[12] Recording still needs a decision, so the agent opens with a consent message. Any mention of severe pain or swelling triggers a warm transfer with a summary. The clinic tests turn-taking with patients who pause while checking a diary.

On cost, the published hosting fee alone for 2,000 minutes a month would be $100, before transcriber, model, voice and carrier charges, which depend on the providers chosen.[9] Our ecommerce product-question agent tutorial shows the same scoping discipline. For recorded messages rather than live calls, see our AI voiceover workflow with review.

Questions readers commonly ask

Is Vapi a voice model company?

No. It orchestrates other providers' models. Our ElevenLabs profile covers one voice provider on Vapi's list.[1]

Does the per-minute fee cover everything?

No. The $0.05 hosting fee excludes transcriber, model, voice and carrier costs.[9]

Can I call customers with a Vapi agent?

Technically yes, with an imported number. In the US, you need prior express consent, and written consent for telemarketing.[8][11]

Is the advertised latency what callers hear?

Not necessarily. Vapi says its displayed total leaves out endpointing and transport.[6]

The next step

Before an agent speaks for your business, settle who answers for what it says. Read who is accountable when an AI agent acts.

About these sources

All pages were accessed on 16 October 2026. Company figures are Vapi's own claims unless stated otherwise. Legal sources are cited for what they state, not as legal advice. The clinic example and figure are illustrative; AI Chronicle has not built or tested a Vapi agent.

Related reading

• LangChain's role in AI agent development
• ElevenLabs and the AI voice business