Beyond Synthetic Consumers: Building a Human + AI Model for Better Brand Decisions

Across our qualitative work in India and APAC, one recurring concern we have heard from clients is that qualitative research can feel too slow and expensive for fast moving decisions. A brand team facing a competitor launch, or finding the winning creative among 3 routes doesn't want to spend the 3-4 weeks that a properly fielded qualitative study requires. 

This gap — between the pace of brand decision making and deep qualitative research outputs — is the issue we wanted to solve without compromising on the depth that a rich consumer study provides.

This is not a theoretical survey of the synthetic-data literature, though we draw on it. It is an account of what we tried, what broke, what held up, and the framework we now use to combine human fieldwork with AI simulation on live client work.

The question we asked was- “How can AI extend the life & value of great qualitative research and help address tactical business questions leveraging this foundational knowledge?’’

Let’s take a step back and start from the very beginning:

What Are Synthetic Consumers and How Are They Created?

“Synthetic consumers” is the industry’s shorthand for AI-generated respondents—typically LLMs prompted or fine-tuned to represent a target population and simulate survey responses or qualitative conversations. The appeal is straightforward: no recruitment, no fieldwork calendar, no incentive budget, and turnaround measured in hours rather than weeks.

Recent research has fueled growing interest in this approach. For e.g. Harvard Business School researchers James Brand, Ayelet Israeli, and David Ngwe demonstrated that LLM-generated preferences could approximate human responses for familiar product categories and can be positioned to rapidly screen ideas before investing in more expensive primary research.

We evaluated multiple AI-enabled qualitative synthetic-data platforms built for online focus groups at scale, AI-moderated interviewing, and synthetic panel generation. Almost all were built and validated against US & Western European data. Almost none held up to the rich, culturally specific landscape of a market like India, where category understanding, language, and decision logic vary sharply not just from Western benchmarks but from state to state and income tier to income tier within the country itself.

What Broke, and What We Learned From It

The turning point for us was not concluding that AI didn't work. It was reframing the question from "what can't AI do" to "what does AI actually do well, and what should stay human." That reframe produced three hard-won lessons that now govern how we use synthetic methods.

1. Garbage in, garbage out is not a cliché, it is the entire methodology.

Credible synthetic consumers depend on relevant, representative, deep, and traceable human data. Research on persona generation similarly shows that synthetic respondents can drift from the populations they are meant to represent unless deliberately anchored in strong underlying data. A polished persona built on thin or unrepresentative inputs can therefore be especially misleading.

2. Conversational elicitation beats numerical elicitation.

Asking LLMs for Likert ratings can produce plausible but shallow numbers. In our experience, richer insight comes from letting synthetic respondents explain reactions in their own words—why a claim feels credible or where a creative route loses them. Numerical outputs can then be derived through structured coding or semantic mapping.

3. Cohort accuracy is real; individual accuracy is hard to achieve.

Synthetic consumers are better at capturing aggregate patterns and cohort differences than predicting individual responses, minority views, or the full variation within a group. Research suggests LLMs tend to gravitate toward socially typical answers, understating the tails. Their value, therefore, is not in predicting what one person will do, but in showing how a cohort’s centre of gravity shifts under different stimuli.

Where We Landed: Human + AI, Not Human vs. AI

Our experimentation didn’t lead us to choose between humans and AI, but to define the right role for each. Human expertise provides the context, research depth, cultural nuance and strategic judgment; AI adds consistency and scale and enables interactive simulations that showcase how consumer decision patterns may respond to new stimuli.

We formalized this into—a two-layer Human + AI model. The foundational human layer uses deep primary qualitative research & expert interpretation to uncover Category Entry Points (the demand triggers that bring consumers into a category) and Decision Archetypes (how people actually make decisions, not just who they are). Depth comes from analysing consumer responses across four layers—declarative, narrative, affective and structural signals. The AI layer then builds on these grounded insights to create Synthetic Decision Archetypes: interactive, simulatable models that can be used to explore new ideas, communication routes, product concepts, or competitive scenarios for up to 12–18 months after the original study, until meaningful market shifts require fresh research.

Simply put: Traditional personas describe people=> Decision Archetypes explain decisions => Synthetic Decision Archetypes simulate Decisions.

Building a usable synthetic twin requires more than demographics or first level psychographics. It needs a structured understanding of decision logic, priorities, frictions, trade-offs, value perceptions, trusted information sources, brand cues and likely responses to different stimuli.

For example, an e-commerce grocery platform was entering a highly competitive, price-sensitive market, the primary qualitative research aimed to understand how lower-income urban households made grocery decisions. Through in-depth interviews, we uncovered a dominant Decision Archetype—Budget Maximizer: 30-to-45-year-olds, small shopkeepers and gig workers, on roughly ₹20-40k a month, whose mindset was "progress within limits" — incremental upgrades rather than aspirational leaps, heavy reliance on social proof, and a self-identity as a wise shopper rather than a discount hunter. That archetype was the foundation for the interactable digital twin that was leveraged months later to identify a preferred creative route without additional primary research.

Scaling Multiple Studies to Cohort Level Insights

Layering above single-study use is can be a Cohort Meta Synthesis: pooling Synthetic Decision Archetypes across multiple studies, cohorts, and markets to unearth shared human truths, emerging motivations, diverging cultural tension points to identify growth opportunities by markets or cohorts.

For example, in a multi-market energy drink study across India, Pakistan, and Thailand, we synthesized Decision Archetypes across two youth cohorts—18–24-year-old college students and 25–35-year-old working professionals—to identify both shared human truths and cohort-specific differences. Rather than producing isolated reports, this framework transformed multiple studies into cumulative organizational learning that can inform future innovation, positioning, and market strategy.

Limitations & How We Guard Against It

Synthetic Decision Archetypes are scoped to the category they were built for; they don't generalize to adjacent categories without fresh grounding. Over-idealized personas that never express doubt or internal conflict are a warning sign, not a feature — we deliberately stress-test synthetic archetypes with edge-case stimuli to see whether they show friction at all.

Inputs need periodic refreshing as category moves, and outputs should be checked against human data and variance, not taken as ground truth on the first pass.

We suggest use synthetic consumers for quick, exploratory work and lower-risk decisions — an early creative screen, a directional read on a positioning idea — and we do not use them as the sole basis for a launch decision, a major investment, or anything where the cost of being wrong exceeds the cost of the primary research we skipped.

Building organizational trust with this methodology has meant checking data provenance on anything we build on, and running anchor comparisons against human data whenever the stakes justify it.

This caution is consistent with the broader evidence. Research from Harvard Business School suggests synthetic methods can estimate average preferences reasonably well, but are less reliable at capturing segmentation, minority views and real-world variation. Across studies, the same principle keeps resurfacing: the stronger and more representative the human data underneath a synthetic model, the more useful and trustworthy its outputs become.

Potential uses for Human + AI approaches	Higher-risk uses that still require direct human evidence
✓ Early-stage concept iteration and refinement	✗ Final go/no-go launch decisions
✓ Message and communication testing	✗ Market sizing or demand forecasting
✓ Exploring multiple creative or innovation routes	✗ Completely new categories with no prior human research
✓ Scenario planning ("What if we changed…?")	✗ High-stakes strategic investments where the cost of error is significant
✓ Internal hypothesis generation before primary research	✗ Regulatory, compliance, or policy-sensitive decisions
✓ Extending the value of existing qualitative research	✗ Predicting individual consumer behaviour or rare/minority viewpoints
✓ Understanding likely cohort-level reactions and trade-offs	✗ Replacing foundational exploratory qualitative research
✓ Cross-cohort and cross-market learning through Cohort Meta Synthesis	✗ Using outdated synthetic archetypes without periodic refresh and validation

What We'd Tell Another Brand or Research Team

1. Treat synthetic consumers as a simulation layer built on primary research, not a substitute for it.
2. Match the tool to the stakes: use synthetic simulation for exploratory work, creative screening and message testing; use human research for foundational or high-risk decisions.
3. Use conversation, not just scores, when you want nuance.
4. Build for cohorts, not individuals; be cautious of claims of individual-level predictive precision.
5. Treat synthetic archetypes as time-bound assets that need refreshing as markets evolve.

Our experience suggests that AI can extend the value of qualitative research when it is grounded in strong human research. The human understanding still has to come first. AI can dramatically extend the value of qualitative research—but only when it is grounded in real human understanding. The human truth still has to be uncovered first.

References and further reading:

• Brand, J., Israeli, A., & Ngwe, D. (2024). Using LLMs for Market Research. Harvard Business School Working Paper 23-062.
• Brand, J., & Israeli, A. (2025). Using Gen AI for Early-Stage Market Research. Harvard Business Review.
• Korst, J., Puntoni, S., & Toubia, O. (2025). How Gen AI Is Transforming Market Research. Harvard Business Review, May–June 2025.
• Peng et al. (2025). Twin-2K-500: A Data Set for Building Digital Twins of Over 2,000 People Based on Their Answers to Over 500 Questions. Marketing Science, 44(6).
• Rahimzadeh, V., Moosavi Monazzah, E., Pilehvar, M. T., & Yaghoobzadeh, Y. (2025). Synthia: Scalable Grounded Persona Generation from Social Media Data. arXiv:2507.14922.
• Hu, Z., Xiao, Z., Xiong, M., Lei, Y., Wang, T., Lian, J., Ding, K., Xiao, Z., Yuan, N. J., & Xie, X. (2025). Population-Aligned Persona Generation for LLM-based Social Simulation.
• Maier, B. F., Aslak, U., Fiaschi, L., Rismal, N., Fletcher, K., Luhmann, C. C., Dow, R., Pappas, K., & Wiecki, T. V. (2025). LLMs Reproduce Human Purchase Intent via Semantic Similarity Elicitation of Likert Ratings. arXiv:2510.08338.
• Bisbee, J., Clinton, J. D., Dorff, C., Kenkel, B., & Larson, J. M. (2024). Synthetic Replacements for Human Survey Data? The Perils of Large Language Models. Political Analysis, 32(4), 401–416.
• Yeykelis, L., Pichai, K., Cummings, J. J., & Reeves, B. (2024). Using Large Language Models to Create AI Personas for Replication, Generalization and Prediction of Media Effects: An Empirical Test of 133 Published Experimental Research Findings. arXiv:2408.16073.
• Gui, G., & Toubia, O. (2023). The Challenge of Using LLMs to Simulate Human Behavior: A Causal Inference Perspective. arXiv:2312.15524.