Artificial intelligence is learning to speak in the voice of the customer. Not merely to answer questions or recommend products, but to simulate preferences, anticipate purchasing decisions, and predict how entire populations might respond to things that do not yet exist. For businesses, the promise is extraordinary: an opportunity to understand markets before entering them, test ideas before investing in them, and reduce uncertainty before making consequential decisions.
But there is a fundamental question beneath this technological ambition: when a machine convincingly reproduces human behaviour, does it actually understand anything about human choice?
For Dr. Anne Scherer, the distinction between plausible answers and reliable predictions is central to understanding what artificial intelligence can, and cannot, contribute to decision-making.
Scherer is a consumer psychologist, researcher, and co-founder of the Swiss company Delta Labs, where behavioural science meets AI-powered customer simulation. Previously an Assistant Professor of Quantitative Marketing at the University of Zurich, she has also conducted research at ETH Zurich and served on the World Economic Forum’s Global Future Council. Her academic work has explored the psychological consequences of increasingly automated relationships between people and companies, including how humans respond to algorithmic advice and machine-mediated interactions.
In her widely viewed TEDx talk, Why We’re More Honest with Machines Than People, Scherer examined a revealing paradox of our technological lives: people can sometimes disclose more to machines precisely because machines do not appear to judge them in the same way other people might. It is an observation with profound implications for an economy increasingly organised around extracting, interpreting, and acting upon signals of human intention.
At Delta Labs, that interest in human behaviour has evolved into a more ambitious undertaking. The company develops technology to simulate how different customer populations might respond to products, pricing, communications, and strategic decisions. Its approach seeks to combine generative AI with empirical behavioural evidence, moving beyond artificial conversations towards simulations that can be tested against real-world outcomes.
Yet this emerging field introduces questions that extend far beyond marketing research. What happens when organisations begin making decisions based on synthetic representations of customers? How do we distinguish statistical confidence from genuine understanding? And what becomes of the unexpected, irrational, or genuinely original dimensions of human behaviour when companies increasingly rely on predictive systems?
These questions become even more urgent as AI agents begin to act as intermediaries between consumers and the marketplace. Brands may soon need to convince not only the people they hope to reach, but also the intelligent systems filtering choices on their behalf. The relationship between desire, recommendation, and purchase could be fundamentally reorganised.
As part of Building Creative Machines’ coverage of Web Summit 2026 in Lisbon, where Scherer participated as a speaker, I invited her to reflect about human honesty, synthetic customers, predictive accuracy, autonomous purchasing, and the future of innovation.
Her answers reveal an essential tension in the emerging relationship between artificial intelligence and human creativity: the better machines become at predicting what already exists, the more important it may become to preserve our ability to imagine what does not.
Image credits: Anne Scherer
The Interview
You’ve explored why people are more honest with machines than with other people. What happens when companies start learning from that honesty?
Companies might finally learn to listen to what customers actually think, rather than what they believe they’re supposed to say.
Human conversations come with social expectations. We want to appear reasonable, responsible, or consistent, even when our actual decisions are driven by emotions, habits, and contradictions. AI can create a space where people feel less judged and are therefore more willing to share uncomfortable truths.
For companies, that’s an enormous opportunity. But the real value isn’t simply collecting more honest answers. It’s understanding the gap between what people say, what they feel, and what they actually do.
And with that opportunity comes responsibility. The goal shouldn’t be to exploit people’s vulnerabilities, but to build products and experiences that genuinely serve them better.
When an AI customer sounds convincingly human, how do you know its answers reflect reality rather than a convincing guess?
A convincing conversation is not evidence of a convincing simulation.
That’s probably the most important distinction in our field. Large language models are exceptionally good at generating plausible human responses. But plausibility isn’t the same as accuracy.
At Delta Labs, we approach this as a scientific validation problem. Our simulations are grounded in empirical population data, including demographic, attitudinal, and behavioral relationships. We then test whether simulated populations reproduce patterns observed in real people, including how preferences vary across segments and how people respond to different decisions.
Crucially, we don’t just compare what people say. Wherever possible, we also validate against what they actually do, including real purchasing behavior.
The goal isn’t to create an AI that sounds like a customer. It’s to create a simulation that behaves like a real customer population, within measurable limits.
What have real customers done that your AI simulations failed to predict?
I don’t have a specific client case I can share here, but one of the most important lessons from our validation work is that human behavior is much more context-dependent than we often assume.
People can express a strong preference for a product or concept, yet behave differently when faced with real prices, competing alternatives, or the friction of an actual purchasing decision. This gap between stated preferences and observed behavior is exactly why we’ve increasingly focused on behavioral validation.
Simulations also have limitations when circumstances change or important contextual information is missing. A model that captures existing preferences well won’t necessarily anticipate an unexpected market shock. But we can introduce such scenarios and different decision contexts into simulations and explore how different customer groups might respond.
For me, the interesting question isn’t whether AI simulations can be wrong. Of course they can. It’s whether we can measure where they’re wrong, understand why, and systematically improve them.
That’s how we move from impressive AI demonstrations to technology companies can actually trust.
When AI agents start making purchasing decisions for us, who will brands really need to persuade: humans or machines?
Both. But they’ll have to persuade them in fundamentally different ways.
Humans buy through a mixture of emotion, identity, trust, habit, and rational evaluation. AI agents will increasingly optimize against explicit preferences, constraints, and goals.
Imagine choosing a hotel. You might fall in love with its atmosphere or the story behind the brand. Your AI agent might prioritize location, price, reviews, cancellation conditions, and your previous preferences.
This creates an interesting challenge for brands. Being emotionally desirable may no longer be enough if an AI agent doesn’t even include you in the consideration set.
But I don’t believe machines will make human preferences irrelevant. Quite the opposite: the brands that succeed will be those that understand what people genuinely value and can make that value recognizable to both humans and their AI representatives.
Could companies become so good at predicting what customers want that they stop creating things nobody knew they wanted?
Absolutely. And that’s one of the biggest risks of using AI the wrong way.
If we only optimize for what customers already want, we’ll get better at creating more of the same. The biggest innovations, from the internet to AI agents, often emerge before people even know they need them.
But that’s where AI gets interesting. It can help us imagine more possibilities AND explore how people might respond to things that don’t yet exist. Because we don’t just model existing preferences. We can model fundamental human needs, motivations, and behaviors, and explore how they play out in entirely new contexts.
All of this is to say that AI should make companies more courageous in their creativity, not more conservative.
The future of innovation isn’t about AI telling us what to build. It’s about combining human imagination, AI-powered creativity, and behavioral simulation to explore ideas we might never have considered before.
Beyond Prediction: Preserving the Unexpected
There is a particularly revealing tension running through Anne Scherer’s answers. The same technologies that promise to make businesses more responsive to human needs could also make them increasingly dependent on models of what human beings are already known to want.
This is not simply a question of whether artificial intelligence can predict consumer behaviour. It is a question of what happens when prediction becomes a dominant organising principle of economic decision-making.
For decades, businesses have attempted to reduce uncertainty through surveys, focus groups, historical data, and market research. AI introduces the possibility of conducting simulations at an entirely different scale and speed. Organisations may be able to explore hundreds of alternatives, model different consumer populations, and evaluate potential decisions long before encountering their consequences in the real world.
But greater predictive capability does not necessarily produce greater imagination.
A system trained to recognise established preferences may become exceptionally effective at reinforcing them. If every new idea must first demonstrate its attractiveness to a simulated version of today’s customer, genuinely unfamiliar possibilities risk being rejected before they have an opportunity to change anyone’s mind.
Some of the most consequential technological and cultural transformations have emerged precisely because someone imagined a future that existing preferences could not adequately describe. The value of an invention cannot always be measured against expectations formed in a world where that invention does not yet exist.
Scherer’s perspective offers an important distinction. The purpose of behavioural simulation should not be to eliminate uncertainty, but to make it more intelligible. A credible model is not one that claims to know everything. It is one whose limitations can be measured, challenged, and improved through evidence.
This becomes particularly significant in a marketplace increasingly mediated by autonomous AI agents. If machines begin deciding which products, services, or experiences deserve our attention, the architecture of consumer choice will change. Companies will need to understand not only human motivations, but also how those motivations are translated into the criteria by which machines compare alternatives.
The danger is that this translation may privilege what can be quantified over what can be experienced. Atmosphere, surprise, cultural meaning, emotional resonance, and the appeal of something entirely unfamiliar do not always fit neatly into a system optimised for measurable preferences.
Yet there is also a more compelling possibility: AI might allow us to investigate unfamiliar futures rather than merely reproduce familiar ones. Instead of asking only whether people would purchase a proposed product, we could use simulation to explore how new environments, technologies, and experiences might reshape human needs and behaviours themselves.
That is where the conversation moves beyond customer intelligence and towards a larger question about the future of innovation.
For Building Creative Machines, Scherer’s reflections speak to a defining challenge of our technological moment: how do we use increasingly powerful systems of prediction without allowing the future to become a repetition of the past?
Artificial intelligence may help us understand human decisions with unprecedented sophistication. It may expose contradictions between what people say and what they do, challenge assumptions about consumer behaviour, and reveal possibilities that conventional research would never uncover.
But the value of that intelligence will ultimately depend on what we choose to do with it.
The most interesting future is not necessarily the one that machines can predict most accurately. It may be the one that human imagination, informed but not constrained by artificial intelligence, still has the courage to create.
Perhaps the ultimate measure of intelligent technology is not how successfully it predicts human choices, but how much possibility it leaves open for choices we have yet to imagine.
Follow us on Youtube and LinkedIn.
This interview is part of Building Creative Machines’ editorial coverage of Web Summit 2026, exploring the intersection of Technology × Artificial Intelligence × Culture × Society.


