Episode 231 | July 13, 2026
AI can build anything. It still needs someone to point the way.
AI moves fast, but humans provide direction. UserTesting leaders unpack AI in the loop, agentic AI, and building trustworthy AI experiences.
AI can build anything. It still needs someone to point the way.
Nobody in enterprise software wants to say it out loud, but the honeymoon with AI is ending, and what comes next looks a lot more like a negotiation.
That's the picture that emerged from a recent roundtable on the Insights Unlocked podcast, where host Nathan Isaacs sat down with three UserTesting veterans—Lija Hogan, Amrit Bhachu, and Mike Mace—to talk about what enterprise leaders are actually wrestling with as 2026 heads into its second half.
The conversation wasn't really about AI and human insight in the abstract. It was about the much messier question underneath: now that AI can do almost anything, who decides what it should do?
The bottleneck nobody agrees on
For the past two years, the tech industry's loudest voices have obsessed over a single question: once you've accelerated engineering, what's the next thing slowing you down? Mike has a theory, and it's not the one dominating the podcasts and the influencer feeds.
"The big issue is how do you get customers to sort of change," he said, describing the gap between how fast companies can ship new features and how fast customers can actually absorb them. Everyone is optimizing the factory. Almost nobody is asking whether the customer wants what's coming off the line.
It's a subtle but important reframe. The industry has spent its energy debugging code. Amrit and Lija spent the rest of the episode discussing that the real debugging needs to happen in judgment, governance, and trust—the parts of the system that don't show up in a sprint retro.
ON-DEMAND WEBINAR
Designing with Confidence at Scale: how AI is changing the way UX teams generate and share insight
The end of free money
Lija's contribution to the conversation started with an economic observation that doubles as a warning: the era of what she calls "token maxing" is ending.
"We are running a gun up against the practical limits of the subsidies that the frontier model companies have been providing to all of us to get us adopting this new technology," she said.
Leadership teams built entire strategies on the assumption that AI compute was, functionally, free. It isn't. And as the bill starts arriving, some companies are discovering an uncomfortable truth: in certain use cases, people are simply cheaper.
That doesn't mean AI is failing. It means the free-trial period of the AI era is closing, and the decisions that come next—what to automate, what to keep human, which model earns its keep—will be made with real budgets instead of borrowed enthusiasm.
Human in the loop, or AI in the loop?
If there's a single phrase this episode pushes back on, it's "human in the loop." Lija doesn't think the industry has the framing right.
"I think we have that term flipped," she said. "It needs to be AI in the loop, because the person there should be a person in the driver's seat."
Human in the loop, in her view, quietly implies that people are passengers—present, but not really steering. And there's a reason that matters beyond semantics: accountability. AI, as Lija put it bluntly, "will never be accountable." Somebody has to be, which means somebody has to actually be driving.
This isn't a small distinction. Swap the words, and you change who gets blamed when things go wrong, and who gets credit when they go right.
Speed versus velocity
Amrit brought a physics lesson to the conversation that's stuck with listeners for good reason: speed and velocity are not the same thing.
"The difference between speed and velocity is direction," he said. "Ultimately, if we're running a really fast pace in the wrong direction, it doesn't matter." AI, in his framing, is an engine for speed. It is not, by itself, a compass. Humans supply the direction—the judgment about which problems are worth solving and which features customers actually asked for.
He illustrated the stakes with a small, sharp story: an airline that canceled his reserved extra-legroom seating, then offered no way to fix it except paying more. A perfectly efficient system, delivering a perfectly bad outcome. "The human touch is really important" in exactly those moments, he said—the ones a script can't anticipate.
Are you an insight seeker?
Join UX, research, and design leaders in London to push your craft further.
Who does the customer actually talk to?
Mike raised a question that deserves more attention than it's gotten: as agentic AI matures and tools like MCP let one AI system talk directly to another, will customers even interact with your brand at all? Or will they send their own AI assistant—their "info butler," as he called his—to negotiate with yours?
"Will customers even come to your bot in the future, or are they going to work through their own bot talking to your bot?" he asked. It's a genuinely open question, and Mike admitted he doesn't know the answer. But he's confident about one implication: companies now need to understand not just how customers use AI, but how they feel about using it, and how they feel about a brand's AI standing in for a human.
That extends to a testing gap Mike sees almost everywhere he looks. Most AI evaluation today measures correctness—did the model give the right answer. Almost none of it measures personality, tone, or trust, even though those are the qualities that will determine whether customers actually adopt an AI experience. "That urgently needs to change," he said, "because that's what will drive customer adoption."
The skills nobody can skip
Underneath all of it sits a demographic problem nobody has solved: if AI absorbs the entry-level work that used to train junior researchers and designers, who develops the judgment this whole conversation depends on? Amrit named it directly, wondering aloud who takes over "as the likes of us get to that kind of point" of retirement, if the traditional path into the field has quietly disappeared.
It's the same tension running through the entire episode, just viewed from a different angle. Democratizing research, scaling insight, letting AI build faster—all of it is only as good as the taste and accountability sitting on top of it. And taste isn't something you download. It's something you build, slowly, the old-fashioned way.
Mike put the larger stakes about as plainly as anyone did all episode: "If everybody can build everything, then it's how do you make that humane? How do you keep in touch with the human beings? Because that becomes the differentiator that's hard for anyone else to provide, and that becomes a basis of lasting value if you can figure out how to build that right."
Episode links and resources
- Lija Hogan on LinkedIn
- Amrit Bhachu on LinkedIn
- Mike Mace on LinkedIn
- Nathan Isaacs on LinkedIn
- Webinar: The 2026 experience survival guide: scaling human insight across every team (UserTesting)—explores how AI and integrated workflows help teams scale human insight across UX, product, design, and marketing, echoing the episode's discussion of democratization and where AI should (and shouldn't) take on more of the work.
- Guide: The early adopter's guide to AI moderation in UX research (User Interviews)—a practical guide to using AI to moderate interviews while keeping a human in the loop, directly relevant to Lija's point about flipping that phrase to "AI in the loop" and to the panel's broader concern about accountability and judgment.
- Podcast episode: The future of customer insights at UserTesting in 2026 (Insights Unlocked)—Mike McDowell previews AI-powered platform capabilities designed to amplify, not replace, human judgment, a theme that runs throughout Mike Mace's and Lija's comments in this episode.
- Podcast episode: Why AI in user research isn't replacing real people (yet) (Insights Unlocked)—makes the case for keeping the human in the loop and using AI for acceleration rather than a substitute for real participants, reinforcing the episode's core tension between AI speed and human accountability.
UPCOMING WEBINAR
