By: Matt Emma
The Fini CEO on why AI customer service didn’t eliminate the labor of running support, and why he thinks most buyers never noticed it just moved.
Did the AI actually stop needing you?
It’s the question Deepak Singla has built his company around answering. And it’s the question, he argues, most of the AI customer service industry has spent the last few years quietly avoiding.
The AI customer service market is on pace to nearly quadruple this decade, from roughly $12 billion in 2024 to close to $48 billion by 2030, according to MarketsandMarkets. Singla’s argument is that a meaningful share of that spending is buying something companies were never told they’d need.
“You didn’t buy a chatbot,” Singla says. “You hired a team to babysit one.”
The Expiration Date
Singla co-founded Fini in 2022, before generative AI had fully entered the mainstream conversation, after watching a support operation at Uber absorb more than 200 million queries a year. What stayed with him wasn’t the volume. It was how much invisible human effort it took to keep the answers accurate as the company kept changing underneath them.
“Every AI is accurate the day it launches, and wrong six months later, because the company moved and the AI didn’t,” Singla says. “Everyone’s fix for that was people. And answering with people reimagines nothing. It just moves the old work somewhere less visible.”
That’s the bet he says Fini was built on from day one: that the technology itself has to learn, rather than depend on a team to keep teaching it.
Not What You Bought
Ask Singla what he’d tell a CX leader who’s already been burned by an AI vendor’s promises, and he agrees with the skepticism outright.
“Part of it is that they were sold the wrong category entirely,” he says. “A chatbot answers questions, and that’s the least interesting thing this technology does.”
What he argues actually exists now is closer to a reasoning system, one that can trace what happened to a specific customer across a company’s internal tools and fix it, rather than just explain a policy back to them. A customer writes in about a failed payment, in his example, and instead of reciting a refund policy, the system traces the transaction, finds the issue, and processes the correction.
The deeper problem, in his telling, is that most systems were built to be at their best on the day they launched, and never again after that. Capability was never really the issue.
“Every AI they’ve bought was at its best on the day of the demo,” Singla says. “Then the company kept moving, the way companies do, and the AI stood still. Six months later they were managing an employee who never learns, one who remembers nothing they were taught.”
Built To Need You
Singla’s sharpest claim skips past any single competitor and goes straight at the incentives underneath the category as a whole.
“There’s an old pattern in enterprise software,” he says. “Sell the tool, then sell the labor to run the tool. AI was supposed to break that pattern. Instead, most of the industry rebuilt it with better branding.”
He calls it inertia dressed up as success. The bot handles the conversations, a new team handles the bot, and the maintenance function gets a “proud name,” in his words, forward-deployed engineering, white-glove service, without anyone asking why a fully autonomous system still needs one.
“Nobody inside the system is rewarded for making the system unnecessary,” he says.
The industry’s own numbers back up how stuck that leaves most deployments. Gartner is projecting that more than 40% of agentic AI projects will be canceled by the end of 2027, citing escalating costs, unclear business value, and inadequate risk controls. Deloitte’s latest Tech Trends research found only 11% of organizations actually have AI agents running in production, despite 38% piloting them.
His proposed test for cutting through it is simple: ask any vendor, including Fini, for the log of every change the system made to itself over the past month, and look at whose name is attached to each entry.
“If every entry traces back to a human, you’re buying a very expensive form you’ll be filling out for years,” he says.
The Quiet Change
Singla’s strongest evidence is less a metric than a description of what a support team’s job looks like once that maintenance work disappears.
“The first change is quiet, and it takes a few weeks to notice,” he says. “The maintenance treadmill stops.”
That job, in his account, was thankless in a specific way: the better someone did it, the less visible their work became. Nobody gets promoted for preventing an embarrassment that never happened. The expertise required to catch a mistake stays the same. What’s supposed to disappear is the requirement that a person keep applying it, over and over, by hand.
“For the first time, a CX team has every conversation the company has ever had, structured and searchable, with a reasoning engine sitting on top of it,” Singla says. “Support turns into where the company learns what its customers are trying to tell it, and answering questions becomes the smallest part of the job.”
Tools You Operate, Infrastructure You Trust
Singla sums up where he thinks this is headed in a phrase he uses often internally: there are tools you operate, and there’s infrastructure you trust.
“Nobody hires a team to babysit Stripe,” he says. “You set it up, you trust it, you build on top of it. Five years from now, I believe it will sound strange that companies once employed people whose full-time job was keeping an AI current, the way it now sounds strange that companies ran their own email servers.”
Getting there, in his view, runs counter to how most of the industry talks about autonomy. As systems get smarter, he argues for tighter guardrails, not looser ones.
“The tighter the boundaries you define, the more you can safely let the AI do,” he says. “Model intelligence is converging fast, so once every model is smart, the thing that’s actually scarce is trust. And trust goes to connected systems that act within well-designed boundaries and stay correct without being watched.”
He’s willing to say what would prove him wrong.
“If it turns out machines can’t keep themselves current in principle, that every AI, forever, needs a human feeding it the world, then the babysitting model was right, and we bet years on the wrong idea,” he says. “Or if regulators decide autonomy is unacceptable no matter how accountable the system is. I take both seriously. But everything we see in production points one direction.”
For now, the evidence he points to sits inside his own company’s changelog. Deeper into a deployment, he says, fewer of the names attached to each entry are human.
Whether the rest of the industry can say the same is the question Singla is betting the next five years on.
