Interviews and surveys tell one story. Usage, renewal, and churn data tell another. For CX leaders, AI’s highest-value role is finding the insights hidden in that mismatch and acting on them.
Most teams now use AI tools for customer insight in a fairly straightforward way: they take interviews, surveys, and call transcripts and turn them into a summary analysis with key takeaways—a neat way to save time and cover vast amounts of data. Recent work on scaling qualitative research with AI is largely about doing exactly this.
But this approach has a significant flaw that should not be overlooked. A simple synthesis of the existing data only surfaces what’s already in it and depends heavily on how the questions are posed and what the AI is directed to look for. I wouldn’t say such an analysis has no value, but it may omit highly important insights because the chosen angle of analysis may simply miss the crucial point.
Take customer interviews, for example. First of all, people tend to sugarcoat their feedback to avoid upsetting the interviewer, and this can vary a lot depending on the interviewee’s cultural background and personality. Secondly, a person might honestly say they would need a certain feature or sign up for a service, and believe it at the moment, but there is an unavoidable gap between what we think we’d do and what we actually do, especially when it comes to spending time or money.
It has long been known in human behavior research that the intention-behavior gap is real: people are quite bad at predicting their own future behavior. Research on this subject suggests that our intentions explain only about a quarter of what we actually end up doing (Sheeran and Webb, 2016).
A famous study on gym memberships by economists Stefano DellaVigna and Ulrike Malmendier showed that people bought monthly memberships, expecting to go to the gym several times a week. Yet, in reality, they went on average four times a month, ending up paying more per visit than they would have with a pay-as-you-go option. They were buying those memberships in good faith, making their estimates for the people they aspired to be, and then life got in the way. As it usually does. And the promises they made to themselves got seriously discounted.
The same thing happens in product usage predictions. An interviewee might promise to use that meditation app every day, pay for that pro feature, or download and test your one-of-a-kind AI app that creates new looks specifically for your body type. But then, to what extent should we rely on those promises if we can’t even keep the ones given to ourselves?
Which is why I propose not just synthesizing what people say but comparing the results against their actions. The gap between those two is where the most valuable customer insights are to be found. And this is where your AI should be pointed. The idea of turning customer input into innovation and not taking it at face value has been around for a while (Anthony Ulwick).
Read customers on two axes
A good way to find real insights is to distribute customer signals along two axes. First, what people say: during customer interviews, in surveys, in sales calls, and in feature requests. Second, what they actually do: usage patterns, conversions, purchases, subscriptions, cancellations, workarounds, or substitute solutions. Plotting these two axes creates a simple 2×2 matrix containing four buckets.
Two of these buckets are clear, consistent, and not very interesting. Real demand: the customer says “yes” and confirms it through actions. This is a great signal. True pass: the customer says “no” and doesn’t do anything. This is simply not your segment.
Now, the other two buckets are worth paying attention to.
False demand: interviews sound positive; the customer is interested in the product and verbalizes their willingness to pay, but then simply doesn’t convert. This is a frequent pitfall because research shows clear demand that would never translate into real KPIs.
We witnessed a similar story at a large foodtech delivery platform, where I served as Head of Strategy. During customer research, one of the core groups of customers looked like a loyal segment and claimed that they were frequent users and that the app was their go-to service for food delivery. But their actions told us a different story. This customer segment would only order food with a coupon, which translated into a clear loss for the platform and confirmed that the segment was primarily discount-driven and would leave as soon as the discounts were switched off.
This case is important as a demonstration of a practical discrepancy. A loyal customer and a discount-driven customer require very different retention strategies. Treating them similarly results in investing in retention of customers who would only stay for the discount and ultimately growing a customer base that drains money and leaves when KPIs switch from high growth to positive unit economics. When we started treating these segments separately, we were able to learn which retention mechanics make sense for long-term growth and which are only hurting the bottom line.
Hidden need: often the most interesting of all. The customer claims that they don’t need a product or a certain feature, yet creates a workaround, buys a weaker substitute, or keeps doing manually the job that the product could fulfill. Words say “no,” behavior says “yes.” Usually this means that real demand exists, but the customer can’t put it into words, or the company has not framed the product in language the customer recognizes.
Why this is a job for AI
Here, AI can be especially useful, and not just for efficient synthesis of thousands of customer data points. Synthesis on its own is helpful, but it smooths over discrepancies. We want to highlight them instead of measuring the average temperature in the hospital. The most time-consuming work here is cross-referencing. We need to take customer interviews, surveys, usage data, KPIs, and support history, and compare them at the level of individual customers or narrow customer segments.
The human mind isn’t capable of keeping the content of hundreds of customer interviews together with millions of usage data points, let alone cross-referencing them by segments. This is where AI steps in and helps compare pages of qualitative information to gigabytes of numerical datapoints, and surfaces hidden insights about inconsistencies between what customers say and what they do.
We saw a good example of this at a B2B AI SaaS company I worked with. Clients were saying that they needed faster and cheaper analysis to help them launch new business verticals, so the product and its positioning were built around this need. At first glance, everything looked consistent: clients asked for speed and low cost; hence, this is what we needed to sell. But their behavior showed that speed wasn’t a real bottleneck for most clients. We delivered fast analysis, but processes would stall after that. There were multiple discussions of the analysis takeaways, numerous stakeholders involved in the approval process for the final decision, and internal politics and corporate dynamics interfering with turning decisions into action.
Looking at the actual behavior of the clients, we could see that what they really needed wasn’t just delivery of fast analysis, but a transparent decision-making framework, defensible to all stakeholders. This translated into a different packaging of the same product, one that addressed clients’ real hidden need.
This explains why so many AI initiatives look promising but often do not translate into real business results. McKinsey’s 2025 State of AI report says that about nine out of ten organizations already use AI in at least one function, yet only 39 percent reported any impact on enterprise EBIT. In my opinion, the problem is that AI is often used as a fashionable tool that delivers impressive-looking output, instead of focusing on applying it in a way that improves the usefulness of the output.
Rank the gaps, then act on the why
It is crucial to note that finding the gap is not enough. The fact that a number of customers said one thing and did the other may just be white noise within the margin of error. But if the same gap is spotted consistently across a significant customer segment and can be translated into real KPIs, it becomes an opportunity worth exploring further. This is especially critical when it comes to the product value proposition. Companies tend to build it on the insights from customer interviews because they can be neatly translated into investment decks. But customer behavior may indicate that the value proposition requires a pivot.
During my work with an early-stage fintech company, I witnessed potential customers getting really excited about the company mission: helping young professionals step onto the path to early retirement by building long-term wealth through developing healthy financial management habits. The product positioning seemed spot on. But customer behavior revealed much more mundane and immediate needs. Customers were asking the AI advisor about ways to raise their credit score, how taxes would be applied in different cases, and how risky a particular investment was.
If we only looked at the behavior, we would come to the conclusion that what’s needed is an AI-powered search tool focused on personal finance. Standalone customer interview takeaways would have produced a financial planner for long-term wealth management. Together, those two signals merged into a more descriptive picture: the larger goal explained why the micro-decisions mattered, and granular questions brought perspective on the mundane challenges that arise on the path to that goal.
The real job the service needed to deliver wasn’t in paving the path to wealth in a general sense. It was in helping customers make the financial decisions in front of them with enough confidence that those decisions would lead them toward the financial future they were trying to secure. This helped us align the value proposition with the real customer need. In the end, this is the job the customer hires the product to do, in Clayton Christensen’s terms.
What to do Monday
You don’t need sophisticated tooling to start. Take one decision that is currently on the company agenda and compare one input on what customers say and one on what they do for the same customer segment. This could be customer interviews and usage data, survey answers and renewal records, feature requests and de facto usage of those features, or sales notes and churn. Upload it into your AI tool with a prompt asking it to synthesize the data, identify gaps across different sources, and formulate hypotheses explaining those discrepancies: Did they claim the importance of the feature but never use it? Did they say that the price was too high yet renew the subscription? Did they say that the product wasn’t needed but create a workaround to accomplish the same job?
Then you can sort those gaps by impact on the bottom line or other product KPIs, take the most important one, and explore what is behind that gap. This would be your starting point to examine your existing product value proposition for how well it addresses the real needs of your customers. Dig deeper into the cases where what customers say isn’t aligned with what they do, and you will find your hidden value, or hidden loss that you can avoid.
Sources
McKinsey, “The State of AI in 2025: Agents, Innovation, and Transformation”
Paschal Sheeran and Thomas Webb, “The Intention-Behavior Gap”, Social and Personality Psychology Compass (2016)
Stefano DellaVigna and Ulrike Malmendier, “Paying Not to Go to the Gym”, American Economic Review (2006)
Harvard Business Review, “How AI Helps Scale Qualitative Customer Research”, Harvard Business Review (2026)
Clayton Christensen et al., “Know Your Customers’ Jobs to Be Done”, Harvard Business Review (2016)
Anthony Ulwick, “Turn Customer Input into Innovation”, Harvard Business Review (2002)