Voice AI Is Quietly Winning the Customer Service War

Call your bank at 9 p.m. about a charge you don't recognize, and there's a decent chance the calm voice that answers has never had a lunch break. It does not sigh when you spell your street name twice. It has read every prior ticket you ever filed. And, increasingly, it closes the case before a human ever picks up. The routine support call, the most human-sounding corner of software, has become the most contested piece of the AI economy.

Four companies tell the story of how that happened, and each answers the same question a different way. Salesforce wants to sell you an agent that already lives inside your customer records. Replicant wants to own the voice call itself. Rasa wants to hand your developers the toolkit and step back. Five9 wants to keep the phone lines it already runs and bolt intelligence on top. Same bet - your next support agent won't be a person - placed from four very different starting lines.

The channel everyone got wrong

For a decade the prediction was that chatbots would swallow customer service. They mostly didn't. The early bots lived in a little window on a help page, answered three questions, and dumped you into a queue the moment anything got hard. What almost nobody said out loud is that the hardest, most expensive channel to automate was also the oldest one: the live voice call. It is real-time. It is unforgiving. A wrong answer to "did my payment go through" is not a bad chat transcript, it is a complaint to a regulator.

That is exactly why voice became the prize. The chat window was a rounding error next to the cost of a phone floor staffed around the clock. Whoever could handle the call - actually resolve it, not just route it - would be selling into the single largest labor line in the customer service budget. Around 2024 the models finally got good enough at holding a multi-turn conversation that the demos stopped being demos.

The last thing in customer service to get automated is the live voice call. It's the hardest, the highest stakes, and now the most contested.

Four bets, one future

The tell in any market is not the marketing. It is the price tag, because pricing is where a company admits what it actually believes it is selling. Look at how these four charge and the strategy falls out of it.

Field guide

How the four approach the same call

Company	The angle	How it charges	Best for
Salesforce	Agentforce, agents living inside the CRM	~$0.10 per action, plus per-user editions	Companies already deep in Salesforce
Replicant	Autonomous voice, built to resolve calls	Tied to automated resolutions	High-volume enterprise phone lines
Rasa	Open-source toolkit you build on	Free core, paid enterprise tiers	Teams that want to own the logic
Five9	Cloud contact center with AI layered on	Hybrid, usage-based virtual agents	Existing call-center operations

Salesforce's pricing history is the most revealing. Agentforce launched at a flat $2 per conversation, hit resistance, and by mid-2025 had been rebuilt around Flex Credits at roughly $0.10 per action, with 100,000 credits selling for about $500. By the end of the year there were per-user editions layered on top. Three or four pricing models in eighteen months is not indecision so much as a company feeling for the seam where a robot's work can be metered like electricity.

Reading the money

Funding and revenue don't crown a winner here, but they do show who is playing which game. Five9 is a public company doing real money at scale. Replicant and Rasa are private bets - one on voice, one on open source - sized accordingly.

By the numbers

Scale of the players (revenue vs. capital raised)

Five9

$1.15B rev

Replicant

~$113M raised

Rasa

~$84M raised

Five9 is the quiet surprise. A cloud phone company most people never think about posted $1.149 billion in FY2025 revenue, up 10%, with enterprise AI revenue up 41% year over year. It launched autonomous Agentic CX voice agents in June 2025 to replace the dreaded press-1-for-billing menu. Its own projection: AI could handle up to 80% of contact-center inquiries by 2029, from roughly 40% today. The boring incumbent turned itself into one of the more interesting AI stories in the market, and did it without a single viral demo.

Salesforce is both a competitor and a backer. Salesforce Ventures put money into Replicant. In this market, the referee also owns a team.

The proof Salesforce ran on itself

The most convincing number in the whole category did not come from a keynote. Salesforce pointed Agentforce at its own help desk and reported that 84% of more than 380,000 support interactions were fully resolved without a human, with only 2% escalating. Running the thing on your own support queue, in public, is the kind of proof that is hard to fake and easy to check. It is also a flex aimed squarely at every buyer who suspects these systems only work in a sandbox.

And yet adoption tells a humbler story. As of May 2025, about 8,000 of Salesforce's 150,000-plus customers were using Agentforce. Price was repeatedly cited as the drag, especially in the mid-market. The demo works. The rollout is slow. That gap - between what the technology can do and what customers have actually bought - is the real state of enterprise AI in 2026, and no amount of press-release math closes it.

The contrarian in the room

Rasa is the odd one out, and deliberately so. Founded in 2016 by Alan Nichol and Alex Weidauer, born in Berlin and now run from San Francisco, it sells the opposite of a black box. Its open-source core has been downloaded more than 10 million times, and a $30 million round in 2024 - backed by Andreessen Horowitz, Accel, StepStone and PayPal Ventures - was raised to fold generative AI into that same developer-first platform.

The Rasa wager is that serious companies, especially banks and healthcare firms, will not rent a conversation engine they cannot inspect. They will want the source, the data, and the logic under their own roof. In a year when most of the industry is racing to sell you the finished agent, betting that the largest buyers want to build their own is genuinely against the grain. It might be wrong. It is not lazy.

So who's winning?

Depends how you count, which is the honest answer nobody in a sales deck will give you. Five9 wins on revenue. Salesforce wins on reach and owns the proof point everyone quotes. Replicant wins on pure voice automation, turning 20 enterprise customers into a reported ~$62 million. Rasa wins on developer mindshare. The market did not consolidate into one victor. It split cleanly down the middle: rent a finished agent, or build your own. Whenever a market splits like that instead of collapsing, it usually means both answers are right for different buyers, and there is room on both sides for years.

For the person whose whole job is answering the phone, the number to sit with is that 80% projection for 2029. It reads as a threat or a promotion depending entirely on which company you work for and what they decide humans are still for. The safe prediction is narrow and probably correct: the routine call - where's my order, did my payment clear, reset my password - is going to a machine. The messy, angry, high-stakes call is staying human for a while yet. The fight now is over exactly where that line falls, and four companies are drawing it in four different places.

Questions people actually ask