How AI Helps Scale Qualitative Customer Research

To make good decisions, companies need information about customer preferences and behavior. The market-research function is thus crucial to market success, but it has an age-old problem: The process of collecting data from consumers is typically hard, slow, and costly. The good news is that generative AI promises to improve this state of affairs. LLMs, in particular, are making it possible to overcome the enduring tradeoffs between breadth versus depth. Quantitative methods (e.g., surveys) offer scale and statistical power but rely heavily on standardized self-reports that provide only a thin view of lived experience. Qualitative research (e.g., in-depth interviews) delivers richer insight into how people think, feel, and decide, but it’s small in scale and ill-suited for generalization. LLM-based interviewers are helping companies overcome this tradeoff. They enable researchers to engage large numbers of respondents—at a scale comparable to quantitative surveys—while capturing much of the nuance, context, and interpretive richness traditionally associated with in-depth interviews. Additionally, LLMs can provide insights that would be very hard, and sometimes even impossible, to generate with traditional approaches. Last summer, we mapped out how gen AI is transforming the collection, creation, and analysis of market insights. That included everything from supporting current practices to replacing them with new methods and synthetic data. Several months later we went deeper, discussing how AI-generated “synthetic personas” and “digital twins” can serve as proxies for real consumers in research and experiments. In this article we’ll address one area where gen AI has already moved beyond pilot programs and academic studies: qualitative research. Companies are now using AI “moderators” to conduct in-depth interviews with real people at a scale and speed that traditional methods can’t achieve. Qualitative research has always been the richest source of customer insight, and the hardest to scale. The relatively high cost, long timelines, and small sample sizes of human-to-human interviews have left business leaders torn: They value the depth of qualitative findings but struggle to act on them with the same confidence they place in structured, statistically significant quantitative data. Managers have long sought an approach that delivers both the contextual richness of an in-depth conversation with the breadth and structure of a survey, ideally at a lower cost. A wave of AI-native startups is looking to finally drive that convergence, and the venture community has taken note. During the past year, for example, Outset, Listen Labs, and Simile each raised $50-$100M from top-tier VCs. Those are just three of the many companies using gen AI not merely to reduce the cost and timing of qualitative research projects, but to make possible research that simply couldn’t happen before. Below, we’ll report on the state of the field and offer a framework to help managers assess and navigate the promising toolkit of AI moderation. When You Need the “Why” Behind the Numbers Perhaps the most significant application of AI-moderated interviews is their ability to blend qualitative depth with quantitative breadth by making it possible to scale qualitative research—an elusive capability that practitioners have long wished for. Microsoft recently demonstrated the power of this capability in its effort to understand consumer perceptions of its products versus those of its competitors. The company’s traditional brand tracker could tell the team what perceptions were shifting across the AI category, but not why. Using Listen Labs to conduct 250-plus interviews with three different audiences, Microsoft piloted what it internally calls “Frontier Listening,” an always-on, semi-structured interview program that captures both open-ended qualitative depth and quantifiable metrics in a single workflow. “This new approach lets us combine depth, scale, and speed in a single workflow, surfacing rich customer nuance in days rather than weeks,” we were told by Rob Graves, a senior director who oversaw the Microsoft project. “By continuously capturing and synthesizing customer perspectives, it turns feedback into actionable insights that guide decisions across teams in real time, while reducing reliance on episodic, slower research cycles.” Recognizing that traditional text boxes often fail to capture the emotional nuance needed for high-fidelity synthetic personas, one of our (Jeremy’s) firms, GBK Collective, partnered with Twinloop, an early-stage company developing AI-driven research methods, to test its AI-moderated voice interviews against standard typed surveys. The results were stark: Verbal responses were, on average, seven times longer than typed ones. The AI’s ability to actively probe and follow up on specific sentiments allowed for a level of “scaled empathy” rarely seen in quantitative research, making it possible to transform flat survey data into rich, multi-dimensional narratives. “Scale” can mean different things in different contexts. Many times, it’s about global coverage. Anthropic, for example, recently announced they had used their Claude-based “Anthropic Interviewer” to conduct more than 80,000 interviews with users in 159 countries and 70 languages, using a combination of predetermined open-ended questions and dynamically produced follow-ups. “AI has enabled us to collect rich, open-ended interviews at an extraordinary scale,” the company reported. “This approach bridges the typical tradeoff in qualitative research between depth and volume.” Sometimes the most valuable insights that emerge from interviews come when human researchers go off script or drill down with more specific questions in real-time. AI-moderated interviews can replicate this dynamic quality at scale. That’s exactly what happened when Sweetgreen worked with Listen Labs to explore what customers wanted from their menu. Listen Labs’ AI discovered that customers wanted not just more of certain macronutrients (such as protein) but also a better ability to see and customize what goes into their meal. Those insights directly informed Sweetgreen’s decision to launch a new in-app macronutrient tracking tool. According to Jonathan Neman, Sweetgreen’s CEO, by using gen AI the company was able to conduct its research at one-third the cost, with five times the responses, and with results turned around roughly five times as fast. In effect, what in the past would have been a multi-week research cycle was reduced to days. When You Need to See What People Can’t Say Most AI-moderated interview platforms are built on large language models and therefore inherently focus on what participants say. But some of the most promising applications capture what people do and feel. Conveo, a Y Combinator graduate, has built AI capabilities to do just that. In a recent collaboration with Unilever, the company conducted AI-enabled mobile-video interviews with consumers in their own kitchens, which allowed Unilever to both capture verbal insights and observe the actual behaviors and attitudes in their natural contexts using multi-modal video insights. The platform then used that data to generate synthesized personas that Unilever’s innovation and insights teams could query interactively to test additional concepts. The process compressed what would typically take months of conventional research into rapid cycles, ultimately producing two product concepts that achieved the highest possible rankings in quantitative validation. When the Topic Is Too Sensitive for a Human Interviewer One of AI’s most surprising advantages in qualitative research has nothing to do with speed or cost. Instead, it’s that people can be more comfortable opening up to an AI interviewer than a human one. When research touches on health conditions, personal insecurities, or other sensitive subjects, it turns out, interviewees often hedge, deflect, or simply don’t show up. A leading men’s health provider discovered this firsthand. An attempt to conduct traditional qualitative research with men experiencing erectile dysfunction, along with their partners, proved to be essentially impossible. Participants wouldn’t schedule interviews, wouldn’t turn up, or would refuse to turn on video. But when the company switched to conducting AI-moderated video interviews, with the help of Outset, participants opened up and provided a depth of insight that the company had no other way to access. Listen Labs encountered a similar dynamic when working with the apparel brand Chubbies, which wanted to talk to young children to understand how to market a new clothing line. They found that children are often more forthcoming with an AI interviewer than with a human stranger. This pattern suggests a counterintuitive principle: In contexts where the personal element of research may create discomfort or social friction, an AI interviewer can often solicit more-valuable insights than a real person. Several recent academic articles have confirmed this conclusion. In one study, for example, when research participants believed they were interacting with a simulated person rather than a real one, they reported less fear of judgment, engaged in less impression management, and disclosed more openly. When Respondents Are Hard to Reach or Schedule Traditional qualitative research assumes that participants can set aside up to an hour or more to engage with an interviewer at a mutually agreeable time. That assumption breaks down with many of the high-value audiences that companies most want to understand. Doximity, an online medical community, faced a version of that challenge. Its target customers—primarily doctors, surgeons, nurses—simply couldn’t block time for live interviews due to their dynamic daily schedules. Using Outset, however, healthcare professionals were able to click a link and complete pre-programmed interviews at their convenience in between patients or late at night. The result was not just more interviews but interviews with people who would probably never have participated in a traditional study. The Road Ahead We see these AI-enabled approaches as dramatically expanding the opportunities for valuable qualitative insight rather than fully replacing the traditional human-led interview. In many cases the two will work well together. Some subjects will continue to require the rapport and trust-building that only a human-to-human connection can generate, such as emotionally nuanced ethnographic work or research where the interviewer’s expertise or credibility is itself part of the methodology (e.g., complex B2B investigations). But for leaders evaluating where AI-moderated qualitative interviews can add immediate value, the cases above suggest a practical framework. AI moderation is still in its infancy. As with any new measurement tool, it will be important to establish key metrics, such as test-retest reliability and external validity. It will also be useful to determine to what extent any gain from AI moderation comes from the modality of interaction (e.g., AI vs. human, voice vs. text) or from the content of the interaction (e.g., the agent’s ability to adjust questions and probe deeper). Academic research into these topics is ongoing. At this stage, before committing to AI moderation, managers would be well-advised to set up a proof of concept, perhaps comparing AI moderation to traditional methods using a small sample. Looking ahead, we expect AI-moderation capabilities to be increasingly combined with the synthetic-persona and digital-twin approaches we explored in our earlier articles. Indeed, companies such as Simile.ai and Twinloop use AI-moderated interviews to collect the data necessary to create digital twins, which are then deployed across various use cases. This depth is critical for training the next generation of digital twins. To rigorously validate this link between deeper data and better prediction, GBK has launched a new study with Columbia Business School and Twinloop to determine what training data mixed with which survey interview modalities yield the most empirically accurate digital twins. Researchers have also started exploring creating synthetic faces for these digital twins—faces designed to look something like what people might look like, given what is known about their personality. Technology is already available to then animate these faces and create digital-twin avatars with whom marketers will be able to conduct qualitative interviews at their convenience—avatars that mimic the thoughts, behaviors, and even expressions of real consumers. When that happens, AI moderation will have made it possible for AI agents to act as marketers who talk to real consumers and as consumers who talk to real marketers. . . . The key benefit of AI moderators is the ability to scale qualitative research. They also allow companies to deploy qualitative research in contexts where it was never previously viable—the quick-turn product decisions, the mid-cycle strategy checks, the market entry where there’s no time or budget for a traditional study. The result will be more qualitative insights that flow into decisions previously made using quantitative data alone or, worse, instinct.

How AI Helps Scale Qualitative Customer Research

Summary.

Partner Center