AI Customer Service Agents Are Eating the Call Center

The last time you contacted a company for help, there is a decent chance the thing that answered was not a person. It knew your order number. It offered a refund without a supervisor. It closed the ticket in ninety seconds and never once put you on hold. Somewhere between the scripted chatbots everyone learned to hate and the AI hype of the last few years, customer service quietly became one of the first jobs software could actually do.

That shift now has a price tag. In January 2026, Decagon, a three-year-old startup that builds AI agents for support, was valued at $4.5 billion after a $250 million Series D led by Coatue Management and Index Ventures. The round roughly tripled its valuation in a matter of months. Decagon is not alone, and it is not even the oldest name in the field. It is one of at least eight companies - Ada, Uniphore, Omilia, Kore.ai, Capacity, Amazon and Five9 among them - fighting over the same unglamorous, enormous prize.

The prizeThe most boring job in the building

Customer service was always the part of a company nobody wanted to fund and nobody wanted to work. High turnover, night shifts, angry callers, metrics like "average handle time" that reduced a human interaction to a stopwatch. Mike Murchison and David Hariri, who founded Ada in Toronto in 2016, did the job themselves at seven different companies before they built a product to replace it. The empathy came first, then the software.

The reason capital is pouring in now is arithmetic. Companies spend a staggering amount answering the same questions over and over - where is my order, cancel my subscription, reset my password. If software handles even a large slice of that volume, the savings are immediate and the market is measured in hundreds of billions of dollars a year. Ada's team has publicly framed customer service as a $750 billion opportunity. Whether or not the exact figure holds, the direction of the arrow is not in dispute.

The shiftFrom chatbot to agent

The old chatbot was a decision tree wearing a friendly avatar. It followed a script, and when your question fell outside the script, it looped you back to the same three answers or dumped you into a queue. What changed is that large language models got good enough to understand what a customer actually wants and, more importantly, to take action - issue the refund, change the flight, update the address, not just describe how you might do it yourself.

Decagon leans on what it calls Agent Operating Procedures, or AOPs, which let a support manager write the logic for an agent in plain English rather than code. That detail matters more than it sounds. It means the people who understand the customer - not the engineering team - own how the agent behaves. Ada takes a similar no-code path. The winning products are the ones a support team can shape without filing a ticket of their own with IT.

Customers don't want to contact support. They want their problem to disappear.The working thesis of the whole category

The fieldInsurgents, incumbents and one voice problem

The market splits roughly into two camps. On one side are the insurgents built for the AI era: Decagon, Ada and Sierra, sales-led and aimed at enterprises with six-figure budgets and months to implement. On the other are platforms that arrived earlier and are retrofitting: Kore.ai, a Gartner Magic Quadrant Leader known for orchestrating many agents inside one governed system; Amazon, folding AI into Connect and its Q assistant; and Five9, the contact-center incumbent adding automation to an existing base of call centers.

Then there is voice, the hardest surface of all. Text is forgiving; a phone call is not. Omilia, positioned as a Leader in Everest Group's 2025 Conversational AI PEAK Assessment, has spent years on voice-driven support. Uniphore, founded in India back in 2008, long predates the current boom and raised $260 million in 2025 at a $2.5 billion valuation - with NVIDIA, AMD, Snowflake and Databricks all writing checks. When the chip makers and the data platforms agree on where AI is heading, it is worth noticing that they landed on customer conversations.

Where the money sits · latest disclosed valuations

Decagon$4.5B

Uniphore$2.5B

Ada~$1.2B

The catchWhat we are not saying out loud

The comfortable version of this story is that AI only takes the easy tickets and frees humans for the hard ones. The honest version is messier. When Decagon lists customers like Chime, Affirm, Oura and Substack, and when airlines, banks and telecoms are the ones writing the biggest checks, the volume being automated is not trivial. Some of those interactions used to be someone's job. The category talks about "deflection" and "resolution rates," and rarely about headcount, because headcount is the part nobody wants printed.

There is also a quieter design idea buried in all of this, and it is the part worth stealing. The best AI support agent is the one you never notice. You do not get a badge telling you a model handled your issue. You just get helped and you move on. Measured that way, the product goal is not to sound human - it is to make the problem vanish so completely that the question of human-or-machine never comes up. That is a very different target than the one chatbots aimed at a decade ago, and it is why this generation is landing where the last one bounced.

What to watchThe next twelve months

Three things will tell you who is winning. First, follow the enterprise budgets, not the demos - airlines, banks and telecoms have the most calls and the least patience, and their contracts are the real signal. Second, watch consolidation: Uniphore alone has absorbed several companies to fill in talent and capability, and the incumbents will keep buying rather than building. Third, watch how honestly each company talks about what happens to support teams. The ones that get ahead of that conversation will keep the enterprise trust they are all competing for.

Nobody grew up dreaming of the call center. It turns out the software replacing it did not either. What is left is a plain question for every company that answers a phone: how much of this can a machine do well enough that the customer stops noticing? A dozen well-funded companies are betting the answer is "most of it," and for the first time the demos are backing them up.

FAQQuestions people are asking