Voice AI spent the last two years proving it could talk. In the third week of September 2026 the industry started arguing about something harder: who pays for it, and on what unit. Three announcements in three days — a per-minute contact center launch, an outcome-priced agent that issues refunds mid-call, and a $21 million raise in mortgage origination — add up to one story. The billing model for voice agents is fragmenting, and the agency that cannot explain the unit economics to a client will lose the budget conversation.
Cheaper to answer, easier to ask
Morgan Stanley published the framing on Wednesday. In a note by analyst Elizabeth Porter, the bank argued that consumer AI assistants — Meta’s Muse is the working example — will expand total interaction volume rather than eliminate it: “Enterprise AI makes it cheaper to answer; consumer AI makes it easier to ask.”
The implication for seat-based contact center vendors is uncomfortable and interesting at once. If clients cut human seats, NICE, Five9 and RingCentral lose the unit their revenue is built on. But if agent-initiated calls and messages rise — authentication, confirmations, follow-ups — usage-based revenue can grow faster than seats shrink. Morgan Stanley named Twilio as the broadest exposure to that shift, Shopify the most direct beneficiary in ecommerce software, and argued Wix is better positioned than GoDaddy for the platforms agents need to transact on.
Treat that as direction of travel rather than a near-term catalyst: the meter replaces the chair.
Three pricing models, all live this week
- Per minute. Voiso launched AI voice agents inside its contact center platform on minutes-based pricing starting at $500 per month with no per-seat fees, aimed at after-hours reception, outbound lead validation and front-line support, with every AI call recorded, transcribed and summarized in the same queues as human teams.
- Per outcome. Yuma extended its ecommerce support agent to voice, billed on tickets it fully resolves. The agent does not just talk: it pulls order history and executes address corrections, cancellations, refunds and subscription updates while the customer is on the line.
- Per seat. Still the dominant legacy unit and the one under pressure from both of the above: seat pricing assumes a human occupies the chair for the whole call, and containment breaks that assumption.
The outcome only counts if the agent can act
Yuma’s launch is the most useful one to study because it names the failure mode directly. “Most voice agents can talk. Very few can actually do anything,” said founder Guillaume Luccisano. “If a customer calls, explains a problem, and the outcome is an email arriving three hours later, you’ve built a slower contact form.”
On a live call the agent identifies the caller, retrieves order history, and applies each change only after confirming it verbally. Every call is recorded and attached to escalated helpdesk tickets. Limited beta now; broader availability planned for October 1, 2026.
The choices worth copying are procedural, not conversational. Yuma built three configurable escalation paths — transfer, email capture with a ticket, or an explicit statement that the agent cannot resolve the request — and lets merchants route a percentage of calls rather than switching over in one step. “Aiming for one hundred percent automation on a phone line is the wrong goal,” Luccisano said. “What matters is the moment a person should take over.”
Where the per-action math gets serious
The strongest signal of the week was Sela’s $21 million across seed and Series A, led by Costanoa Ventures with Emergence Capital participating. Sela runs voice agents for mortgage origination and reports more than $1 billion in originations a month, a $10 million annualized run rate inside 18 months, and six of the ten largest independent mortgage banks using it in production. Those figures are company-reported — treat them as directional, not audited.
The interesting part is the unit of value. “We measure success in borrowers reached, productive conversations held, and loans closed, not minutes used,” said CEO Nate Becker. Costanoa’s David Cheng framed demand as pressure on the two numbers a lender’s P&L cares about: conversion and cost per funded loan. That is the shape of a defensible voice AI product — a narrow, expensive workflow where the agent escalates to a human when the conversation requires one.
When the “AI” has someone on the other end
Reuters reported Tuesday that Meta has been testing a human concierge behind Muse: contractors taking over some of the calls the agent places, with the option enabled for half of Meta’s employees. Internal posts flagged that sensitive information could reach call-center contractors even where nobody intended it to. Meta said rollout would come “with the proper disclosures.”
Note what Meta’s privacy language does not cover: saying each agent runs in its own virtual machine describes where the software executes, not who is listening. Meta has been here before — its M assistant, tested from 2015, was staffed by human trainers and never cleared roughly 30% automation before it was shut down in 2018.
The lesson for agencies is contractual. Automation level is a disclosure item. If a client’s brand voice is answering the phone, that client is entitled to know whether the thing on the line is a model, a person, or a queue of contractors in another time zone.
The latency bar keeps rising
Tencent’s research model Gander shows where caller expectations are heading: a “cerebellum” holds the conversation second by second while a swappable “brain” runs agent work in the background, so the system keeps talking while it works and users can interrupt at any moment. On Full-Duplex-Bench v3 it interrupts users in 8% of scenarios, against 13.5% for GPT-Realtime, though it still trails on task accuracy.
Naturalness has a benchmark now, too. SquadStack tracks Abruptly Disconnected Rate — callers who hang up in the first ten seconds after realizing they are talking to AI — and reports around 10% for its agents, inside the 8–12% range it measures for humans. Whatever you call it, pick a metric that measures the caller’s reaction, not the model’s confidence.
What to put in the client contract
- Bill on the unit you can measure. Minutes for a line that only qualifies and routes; resolved interactions for an agent that can change a record; nothing outcome-based for an agent that can only talk.
- Define “resolved” in writing before launch. Yuma’s definition — the action completed in the system of record — is dispute-proof. “Call handled” is not.
- Instrument the handoff, not just the bot. Containment without handle time, repeat-call rate and CSAT is a vanity number.
- Ship escalation before cleverness. Transfer paths, expected response times, and an honest “I can’t resolve this” are the parts clients worry about. Build them first.
- Disclose the automation level. Written into the statement of work, not discovered by a customer.
What to watch
Two testable questions next quarter: does per-outcome pricing survive its first round of resolution disputes, and does agent-initiated traffic show up in usage-based revenue rather than just replacing seat revenue. The answer decides whether the meter is a real business model or a repackaged seat. Either way, the agencies that win this cycle will be the ones that can name their client’s unit of value before the invoice does.