A startup ran the same question through a synthetic panel and a real consumer sample: “What would you improve about facial masks?” The result? 90% of real respondents said nothing; fine as is. The synthetic panel returned a detailed improvement roadmap. While both insights are valuable, it’s important to recognize that they’re measuring different things.
To understand why that gap exists, you need to understand the three waves of AI reshaping consumer research from the inside out. Across the global CPG and retail companies we work with, we see the first wave already embedded, the second increasingly adopted, and the third being piloted for targeted use cases. Together, these waves are turning consumer research from a centralized function into something much broader.
The Analysis Wave
The first and most established wave is AI for analysis; summarizing transcripts, coding open-ended responses, synthesizing verbatims. This is now commoditizing into a feature across research platforms rather than a standalone value proposition. But getting analysis right remains the foundation. If AI distorts at the interpretation layer, every downstream decision inherits that distortion.
The known biases — sentiment softening, framing drift, primacy effects — are increasingly well understood. Practitioners are addressing them through persistent learning loops and cross-referencing across data types.
Two of the most important breakthroughs are cross-project intelligence and multimodal signal layering. While cross-project intelligence enables the AI to continuously learn across previous studies, multimodality reduces AI’s tendency to flatten emotional content into tidy text summaries, by analyzing not just transcript text but vocal tone, facial expression and gesture simultaneously. As such it is a critical component as AI moves from analyzing existing data to moderating live conversations.
The Moderation Wave
AI-moderated qualitative research has moved from pilot to production. Over $250M in venture capital has entered this category in the past twelve months, backed by VCs such as Sequoia, Lightspeed or Microsoft’s venture fund. The investment thesis is clear: qualitative depth at quantitative scale.
But the most interesting finding around adoption isn’t about speed or cost, it’s about candor. Across our clients that deployed AI moderation tools, a consistent pattern is emerging: respondents open up more to an AI interviewer than to a human one. On sensitive topics, or in contexts where social dynamics constrain honesty, removing the human moderator lets the human interviewee go deeper. Voice-based AI interviews produce responses dramatically richer than typed surveys. This isn’t an efficiency gain; it’s a qualitatively different data source.
As AI moderators still carry tendencies toward sycophancy and uneven probing, frontier vendors leverage prompting and methodological guardrails to mitigate these. Here the multimodal capabilities of the analysis wave become critical, as live probing requires understanding not just what was said but how.
And once thousands of real conversations flow through an AI-moderated platform, a new possibility opens: using that primary data to generate new responses.
The Generation Wave
What happens if, instead of AI only analyzing or moderating interviews, it fully generates them? Whether via synthetic personas, panels or fully simulated populations, this sounds like the epicenter of AI biases. And yet, the generation wave is currently emerging as the third wave of AI disruption in consumer research. What is clear is that the economics are transformative: weeks to hours, six figures to near-zero marginal cost.
But the central challenge is just as obvious to anyone who has ever worked with AI. The training process that makes LLMs helpful and agreeable is the same process that destroys the variance consumer research depends on. Industry leaders call this the distribution problem, and it shows up on two axes.
First, across the panel, synthetic respondents cluster around the most probable answer, erasing the tails (e.g. early adopters, contrarians). Second, within each respondent, the same flattening happens. While real consumers often hold competing views in a single answer, claiming they love a brand while listing three things they hate about it, synthetic personas resolve that tension before they speak.
To solve for this, startups aren’t converging on a single solution. The foundation is what any good researcher would do: A/B test. By running multiple synthetic rounds against a real panel with identical briefings, they can use the delta to build refined correction layers around the LLMs each cycle. The result? A pilot run by a global industrial brand showed 90% response overlap with human data within the first two weeks.
That said, frontier providers are moving past these surface corrections, arguing that the tendency toward the middle is inherently trained into the weights of general-purpose LLMs. Moving beyond correction layers, they increasingly use the accumulated human-synthetic delta as training data to build their own small language models, or SLMs, trained on a narrower, purpose-built dataset.
A beauty SLM, for instance, learns how Canadian women aged 35-55 actually respond about skincare, not how the average internet user writes about it. If the bias sits in the weights, that’s where you fix it. Through this approach, practitioners and researchers alike believe that every hybrid study makes the next synthetic study better.
Others start from a completely different premise: that language models are the wrong foundation for modeling human behavior. Building what might be called large behavioral models, architectures rooted in cognitive science and observed decision-making, they bypass language prediction entirely.
Some startups model psychological mechanisms that drive consumer choices; others, like Unbox.AI, train on transactional signals like clicks, scrolls and purchase sequences. What they share is the conviction that the distribution problem is partly an artifact of modeling what people say rather than what they do. When you build from behavioral data, the messy variance that language models flatten is preserved by design, and the intent-action gap, the oldest unsolved problem in consumer research, becomes structurally narrower.
The Bigger Shift
When analysis is embedded, moderation is scalable and generation is near-free, consumer signals are no longer gated by the CMI team’s budget and bandwidth. As product managers, innovation teams and strategists gain access to consumer signal faster and more frequently than ever before, the CMI team’s role will elevate, from running studies to architecting decision intelligence by democratizing access across the organization, as well as defining when ‘real’ insights are needed. As Nader Fadl, Founder of Experial, puts it, “synthetic where possible, human where necessary.” Not a caveat, but a design principle.
And there’s a quieter reason to move now. The hybrid datasets being built today (e.g. the deltas between synthetic and real responses across categories and audiences) are the training data that tomorrow’s synthetic research will depend on. Whoever waits doesn’t start later at zero; they start at a deficit.
So the question for every retail leader is not whether to adopt AI research. It’s how to build processes that ensure biases are monitored and mitigated, define clear criteria of when to use which level of AI, and support the democratization of consumer research throughout the organization. This is the path in which consumer research becomes decision intelligence.
At Silicon Foundry and previously at Kearney’s IMP³ROVE Innovation Competence Center, Clemens Pfefferkorn advises clients across private and public sectors on innovation strategy, venture engagement and ecosystem development. His work has spanned technology landscaping, venture partnering and institutional design, including leading a global startup scouting for an international energy conglomerate and supporting the setup of innovation structures that connect corporates, startups, investors and policymakers. Earlier, he gained hands-on startup experience as Employee #2 at a sport-sponsoring venture, supporting go-to-market and business development. Pfefferkorn holds an M.Sc. in International Management (CEMS) from the University of Cologne and the University of Cape Town, where he dual majored in Corporate and Social Innovation.
