If you’ve been in a strategy meeting recently, someone has probably mentioned synthetic audiences. The pitch is compelling: AI-generated consumer panels that simulate real buyer behavior, delivering research insights in hours instead of weeks, at a fraction of the cost. Vendors like Qualtrics, Electric Twin, and others are moving fast, and adoption is accelerating. By some estimates, 73% of market researchers have already used synthetic responses at least once.
For certain applications, the technology delivers. Early-stage concept screening, directional testing, expanding a small sample to identify broad patterns. These are legitimate use cases where synthetic data can save time and budget without meaningful risk. If the question is “which of these three taglines resonates more with mid-market IT buyers,” a well-built synthetic panel can get you a reasonable signal quickly.
But in B2B research, particularly win/loss analysis and voice-of-customer studies, the limitations are not minor. They are structural.
The sycophancy problem
AI models are optimized to be helpful and agreeable. In a research context, that tendency becomes a liability. Studies from Nielsen Norman Group have found that synthetic respondents are consistently more positive than real humans. They underreport dissatisfaction, avoid strong negative opinions, and tend to provide feedback that looks encouraging rather than honest. In win/loss research, where the entire purpose is to understand why a deal was lost or a customer nearly churned, an optimistic simulation is worse than no research at all. It validates assumptions instead of challenging them.
The variance problem
Synthetic panels converge toward the average. They are trained on patterns, and patterns produce pattern-typical responses. A comparative study by Emporia Research found that when real B2B respondents were asked about salary satisfaction, answers ranged from “Somewhat unsatisfied” to “Strongly satisfied.” The synthetic panel? 98% answered “Somewhat satisfied.” On work-life balance, real respondents spread across the spectrum. Synthetic respondents chose the same answer 100% of the time. The data lacked the variance that makes research genuinely useful. In voice-of-customer research, the insight that reshapes a product roadmap or retention strategy almost never comes from the median response. It comes from the outlier: the customer who describes a pain point in language the product team has never used, or who reveals a competitive dynamic that internal teams didn’t know existed.
The follow-up question problem
This may be the most important limitation. In a live interview, a skilled analyst hears something unexpected and follows it. The second or third question, the one that wasn’t on the discussion guide, is often where the most valuable insight surfaces. A synthetic respondent can only answer the question you thought to ask. It cannot volunteer what you didn’t know you needed to hear.
None of this means synthetic data has no role. It does. As a complement to primary research, it can extend reach, accelerate early-stage exploration, and help pressure-test hypotheses before committing to a full study. The technology will continue to improve, and responsible organizations should understand where it fits.
But the decisions that shape competitive strategy, customer retention, and product direction in B2B still require a real conversation with a real person. Someone who will tell you the uncomfortable truth that a model trained on patterns cannot simulate.
The question for any organization evaluating synthetic data is straightforward: what is the cost of a confidently wrong answer? In low-stakes, early-stage contexts, that cost is manageable. In the research that drives your most consequential decisions, it is not.
Synthetic data can tell you what’s probable. A well-conducted interview can tell you what’s true.
James Rice – Chief Digital Officer
