Unite AI just ran an interview with Neal Lathia, co-founder and CTO of Gradient Labs, and if you care about where agentic AI is actually going versus where the hype says it's going, it's worth your time. Per the interview, Lathia has nearly two decades of machine learning and data science experience across fintech, consumer platforms, and academic research, and you can feel that background in how the piece frames the problem his company is trying to solve.

You can read the full interview over at Unite AI.

Gradient Labs builds AI agents for customer operations, focused on financial services, and the framing is the part I found genuinely interesting. The hard problem with agents isn't capability. It's reliability. Getting an agent reliable enough that a company will let it talk to real customers without a human babysitting every reply is the whole game. As someone who spends his days shipping software, this rings true. The gap between a demo that works 90 percent of the time and a production system a business will bet its reputation on is enormous, and it's mostly an engineering and evaluation problem, not a model-size problem. Regulation comes up too, which makes sense given the financial services focus. If you're deploying agents in a regulated space, reliability isn't a nice-to-have, it's the entry fee.

That reliability framing is also why the background matters. Someone who spent years in fintech and consumer platforms has presumably lived through what happens when a model prediction goes wrong for a real customer, and that's a different education than the usual founder-on-X thread. I'd take opinions on agent trustworthiness from that kind of resume over the benchmark-chasing crowd any day.

What I keep coming back to is the evaluation question, because it's the same wall my own field hit years ago. Once your system makes autonomous decisions, your testing has to get way more sophisticated than a benchmark score. You need to know what the agent did, why it did it, and what happens when it's wrong. Whether the interview digs into how Gradient Labs handles guardrails and monitoring, I'll let you find out for yourself, but that's the question I'd be asking a company deploying agents into regulated customer conversations...

My own read on where this goes: the likely future isn't AI replacing support teams wholesale. It's AI handling the routine volume so humans can take the cases that need judgment. Less exciting headline than "AI replaces support," sure, but it's the version that matches what I see actually getting deployed. And companies in regulated spaces especially don't get to skip the trust part.

It's also just a good read on the founder perspective in the current AI moment, when every week brings another benchmark leap and another round of "agents are here." The gap between benchmark and deployment is where companies like this live or die, and interviews like this one are one of the few places you hear the problem stated plainly.

And if you've ever had a support agent, human or otherwise, confidently tell you something completely wrong, you already understand the stakes.