We’re not normal: What the AI adoption gap means for the products you’re building

Chances are you used an AI tool before 9 a.m. today. Chances are most of your users didn't.

It's a simple contrast, but Debbie McMahon, VP of Product at Loveholidays, built an entire MXP London session around it and what it means for every product decision. She called it the AI adoption gap, and she was refreshingly clear-eyed about the ways it had distorted her own judgment.

An infographic Debbie shared on stage breaks the world's relationship with AI into four tiers:

That last group is the one that fills rooms like MXP London. And Debbie's point is that this creates a fundamental disconnect. The people designing AI-powered products are a statistical anomaly relative to the people they're designing for.

She made that concrete with a story about her friend Agnes. Agnes is a primary school teacher who spends her days helping children from difficult family circumstances settle into school. She's smart, busy, and has a free ChatGPT account she’s never used. Why? Her school doesn't allow AI tools on work devices. The budget isn't there for anything paid. And nobody on staff is confident enough with the technology to know where to start, even if the first two problems disappeared tomorrow.

Agnes is not an edge case. She's a typical Loveholidays customer, and she's who Debbie asked the MXP London audience to hold in mind.

For a look at how AI engagement is trending across real products, Mixpanel's 2026 AI benchmarks draw on nearly 290 billion AI events across 2.6 billion devices, and the picture is more nuanced than adoption headlines suggest.

Debbie named three patterns she sees in herself and across the industry.

Over-prioritizing AI features. "Have you put something out there just because it was easy to do?" She raised her own hand. The pressure is real: boards want AI in the roadmap, investors want AI in the roadmap, competitors are shipping. And the tooling makes it fast to build something. But fast to build isn't the same as worth building.

Underestimating friction. Most people are forming their views on AI right now without the benefit of product communities or conference sessions to help them process it. That's not a knock on users. They're figuring out whether to trust the technology at the same moment they're being asked to use it. Dropping a disclaimer like "generated by AI, so may not be accurate" into a set of results doesn't help someone who's already uncertain about the output.

Asking users to change behavior. Behavior change is extraordinarily difficult even when people want to change. For users with an ingrained habit and no clear reason to do anything differently, it's close to impossible. The customers who are confused or skeptical about AI aren't going to reroute their entire search behavior for a feature they weren't asking for. If what they want is a keyword search and a filter, that's what they'll reach for.

Two case studies from Loveholidays that Debbie shared with the kind of candor that makes MXP sessions worth attending.

The room grouping problem. Hotel listings at Loveholidays had accumulated a long tail of duplicated, poorly named room types, a messy, years-old problem that had resisted every previous attempt to fix it. When a team went after it with AI, they went in with confidence. What they found was that the problem was deterministic: two rooms are either the same or they're not, and AI isn't built for that kind of precision. The model grouped rooms in ways that didn't hold up to scrutiny, renamed options customers recognized, and quietly dropped listings that people had been booking. The output looked coherent, but it wasn't.

In Debbie’s own words: "Don't put AI on top of a broken thing. Fix the actual problem."

The Q&A tool. Another team shipped an LLM-based Q&A feature into the pre-booking journey, built around the hypothesis that customers comparing hotels would want to ask questions. The feature got used, a lot, actually. But when the team looked at who was using it, the data didn't match the hypothesis at all.

The users weren't pre-booking customers weighing their options. They were people who had already made a booking, navigating a long and confusing post-booking experience to find basic information about their upcoming trip. And the questions they were asking weren't the ones the team had imagined. The top two were whether their room came with a kettle and whether it had an iron. Third most common: the wi-fi password.

In other words, customers who needed information Loveholidays already had were threading through a complex AI-powered feature just to ask if they needed to pack a travel iron. The team proposed moving the tool into the post-booking journey where the actual demand was. Debbie shut it down, and the reasoning is worth sitting with.

The post-booking journey got fixed instead, and the Q&A tool got pulled. The story doesn't end there: the team is relaunching the feature with a genuinely different hypothesis behind it. Since that original experiment, Loveholidays launched its own reviews platform, which changes what a Q&A tool can meaningfully do and who it's actually for. There's a real use case now tied to a specific user group, and that's what "not now" looks like when it's used well.

Loveholidays went to their customers directly and asked how they'd feel about AI-mediated booking experiences. The words that came back included "anxiety," "I trust filters more," and "I prefer clicking." There was also genuine enthusiasm from a portion of respondents. But the skepticism was real, present across a significant share of the audience, and Debbie didn't try to explain it away.

53% of consumers distrust AI-powered search results, according to a Gartner survey published in September 2025. Debbie cited a similar figure on stage and her read was plain: "I can't ignore that. That’s the reality for my customers, and that's where we're starting from."

Her answer is to segment more precisely. AI comfort and interest now sit alongside demographics, holiday type, and referral route as a dimension Debbie's team uses to personalize the experience. Design one AI-mediated journey for everyone and you either hold back the users who want to explore new things or push unwanted change onto the customers who aren't ready for it.

Debbie's dad started using email about a decade after she did. She uses that as a reference point: adoption follows its own timeline, and you can't compress it by shipping a feature and calling it done.

Recognizing a problem clearly is useful. Building an organizational system to address it at scale is something else. That's what separates this MXP session from a general mindset piece.

Debbie's team built what she calls "out the bubble," an internal prototyping environment that anyone at Loveholidays with access to Codex or Claude Code can use, with no prior technical setup required. It's pre-integrated with the company's design system, data mesh, live pricing, and real images, which matters more than it might sound. When you put a prototype in front of a customer in a travel context, the first thing they notice is whether the hotel price looks real. This system removes that distraction so the feedback is about the experience itself.

The pipeline works in three stages:

The important shift isn't the speed, though. It's that the people with the best product ideas are no longer gated by their ability to build prototypes. The system makes that capability available to the whole company. Fewer ideas get duplicated. More of them get in front of real customers. And when something earns the full investment of a real build, there's actual evidence behind the decision.

Debbie closed by revisiting the Henry Ford quote, carefully, because she's not entirely sure Ford said it. But the argument holds regardless. The car achieved adoption because Ford understood what his customers actually needed. He built a faster horse, not a different animal entirely.

The tools have changed. The product job hasn't. What AI does change is how big your answers can be once you've understood the question clearly. The potential is genuinely extraordinary, but only after you've done the harder work of understanding what your users need and why they'd change.

Remember Agnes from the opening of this piece? The primary school teacher who supports children through difficult family circumstances, has a ChatGPT account she never uses, and has no budget, no tech support, and no time to figure it out herself. She's never going to ask for an AI-powered solution to her problems. But she'd benefit enormously if someone took the time to understand those problems and built something that addressed them, with AI as part of the answer rather than the starting point. 

That's what good product work looks like, and as Debbie put it, that's always been the job.

The product teams pulling ahead right now aren't the ones shipping the most AI features. They're the ones who understand the gap between their own relationship with AI and their users', and who've built the systems to keep closing it.

To understand what your users are doing in your product, Mixpanel AI is built to surface that signal proactively, so the insight finds you rather than the other way around. For technical teams who want to go further, try Mixpanel Headless where AI agents read and drive Mixpanel programmatically.