Costs and pricing

How the money moves

The business pays everyone, usually per minute, and the meters start at different moments. Here's the per-minute stack at US list prices in October 2026.

In a mid-priced chained stack at about $0.12 a minute, the platform takes about 45%, the language model and the voice 10-30% each, telephony 5-12% and speech to text about 5%. Model prices fell while platform fees seem to have held at about $0.05: OpenAI cut its realtime audio prices 60% in December 2024 and another 20% in August 2025, and ElevenLabs halved its agent prices in February 2025.

The traps sit in the billing units. AssemblyAI bills streaming by session time, silence included, and GPT-Live-1 bills the whole session, backend thinking included. Telnyx rounds each call up to the minute, so wrong numbers cost a full minute. Speech-to-speech models re-bill the conversation each turn, and changing instructions mid-call breaks the cache. And plans cap simultaneous calls, with extra lines at about $10 a month each at Vapi.

The AI minutes are rarely the big number. I modelled the payroll company's line (100,000 calls a month of about four minutes each) against in-house staff paid the US median wage for customer service work, plus benefits and overhead, and the AI usage came to 4-12% of the labor bill it replaces. Containment sets the savings: moving it from 30% to 60% saved five to ten times more than halving the AI price per minute. Against offshore staff, an agent that contains only 30% of calls can cost more than it saves once the fixed costs (conversation designers, evals, integration) are counted. An agent that only verifies callers and routes them, with people taking every call, saves much less, and that's mostly how contact center incumbents sell AI: lower risk, smaller savings.

Priced per outcome instead, the same line's 60,000 resolutions a month would cost about $60,000-150,000 at the $0.99-2.50 per resolution that Fin lists and others reportedly charge, against $14,000-72,000 of raw AI usage for the same calls. The vendor keeps that spread for carrying the failures and the model costs. I think buyers underrate the outcome definition: Fin bills handoffs that follow a procedure and self-serve routings as outcomes too, and a caller who hangs up angry still counts as contained.

How companies of different sizes use it

Headcount matters less here than call volume and, once there's a contact center, the number of agent seats, because those set the product, the vendor and the bill. Nobody publishes seat counts or who owns voice AI by company size, so the bands, the owners and the spend below are my rough estimates, built from list prices, startup programs and one procurement broker's data.

A startup (under 50 people, and at most tens of thousands of AI minutes a month) comes in two kinds. A voice-native startup building an agent product assembles the developer stack (Vapi, Retell or LiveKit, speech models, a lab's model, numbers from Twilio or Telnyx), and its voice bill is its cost of goods. The CTO or a founding engineer owns it, and the CEO watches the margin per minute. Deepgram and ElevenLabs give startups a year of free credits and LiveKit has a free tier, though Vapi's program had paused new applications in October 2026. An ordinary startup adds voice to its chat support agent, and a dental clinic or a plumber buys an AI receptionist for a flat monthly fee. My rough estimate for a voice-native startup is $7,000-90,000 a year, much of it covered by credits at first. The push to the next step is the credits running out, the free tiers' caps on simultaneous calls, and the first customer that needs a BAA or SOC 2.

A scale-up (about 50 to 500 people, and up to a few hundred thousand minutes a month) splits the same way. A voice-native scale-up negotiates model and telephony rates, and some move to LiveKit or Pipecat to drop the platform fee. A company deploying for its own support usually runs a small team on a help desk such as Zendesk, Intercom or Salesforce, and voice arrives as the help desk's AI add-on, a small finished-agent contract such as Fin's, or an engineer's build on a committed Vapi or Retell plan. The head of support or CX operations owns it, with one engineer on integrations. My rough estimate is $50,000-300,000 a year, more for a voice-native company whose minutes are its cost of goods. What pushes it on is passing a few dozen agents, where contact center vendors' seat minimums start, an enterprise customer asking for a BAA or EU data residency, or the first outbound campaign that brings TCPA consent records into scope.

A mid-market company (about 500 to 2,000 people, or 50 to 300 agent seats) rents a cloud contact center, often with an outsourcer taking overflow, and the payroll company from the start of this guide sits here. It mostly buys voice AI as an add-on to that contact center. Five9 bundles AI minutes with each seat, Genesys meters AI in tokens on top of its seats, and Amazon Connect charges by the minute with AI included and no seats at all. Some put a finished agent such as Decagon or Parloa beside the contact center instead. The VP of customer service owns the outcome, IT owns the contact center contract and the phone lines, and legal reviews recording and disclosure. My rough estimate is $150,000 to $1.7 million a year on voice AI, on top of the seats, and containment decides how much of it comes back. Containment that stops improving, a new language or country, and the contact center's renewal push it to the next step.

An enterprise (about 2,000 to 10,000 people, or 300 to 2,000 seats) runs several contact center sites, in-house and outsourced, in several languages, with regulated flows for payments, health or identity. This is where finished agents sell: Sierra, Decagon, Parloa and PolyAI on annual contracts that run to six figures a year in the broker's data, next to AI from NICE or Genesys, voice security from Pindrop or Nuance, and Twilio as the pipe. A senior customer operations leader owns the program, with the CIO's contact center team, procurement, legal (consent, wiretap, biometric, health and card rules) and security for voice authentication. My rough estimate is $500,000 to $3 million a year, since most programs cover a share of call types rather than every call. The cloud commitment starts to pick the vendor here, as it does in cloud security: Decagon listed on AWS Marketplace in July 2026 so buyers can pay from committed AWS spend. Mergers, new countries, renewals and board pressure on the cost per contact push it further, and in Spain a new law caps how far it can automate (more in Tailwinds and headwinds).

A Fortune 500 company (the largest US companies by revenue) runs everything at once: several contact center platforms after acquisitions, several outsourcers across countries, voice biometrics, and its own builds on labs' models beside finished agents. Sierra says it serves over 40% of the Fortune 50, PolyAI lists Marriott and Caesars Entertainment, and Amazon's Ring chose Vapi's platform, all as the vendors tell it. Some of their voice AI comes through their outsourcers, which now sell AI services as call volumes shrink. A chief customer officer or COO owns it with an AI program office, alongside the CIO for platforms and vendor management for the outsourcers. My rough estimate is $1-5 million a year for a cautious program on a few call types, and several times that for a consumer brand that commits fully and gets containment to hold. There's no next size up; the pressure comes from regulators, outbound lawsuits at scale and deepfake fraud against voice ID.

What changes as a company grows (bands and spend are my rough estimates):