09Multi-Agent Persona Simulations for Early Product Research (Consumer Banking)
Synthetic focus groups stress-tested product concepts before real research spend
Role: Solution Engineer / Prototyper
Executive summary
Prototyped an AI persona framework using multi-agent LLM simulations to generate synthetic customer personas and stress-test new product concepts before costly market research.
- Azure OpenAI
- Microsoft Agent Framework
- TinyTroupe (Microsoft Research)
- NLP sentiment analysis
A consumer-banking client explored using AI for market research and product development—creating customer personas and simulating audience reactions to new financial products before expensive real-world studies. We proposed generative AI and multi-agent systems to create synthetic personas, run simulations and extract insights.
- Use LLM prompts to create detailed, realistic personas (e.g., tech-savvy millennial investor, risk-averse retiree) with attributes and behavioral scripts.
- Place persona agents in scenarios mimicking early product research (e.g., a simulated focus group on a new card feature).
- Capture conversations and decisions to derive insights (likes, objections, questions).
- Evaluate whether simulations yield plausible, valuable insights to supplement traditional research.
I built a sandbox on Azure OpenAI using Microsoft Agent Framework for multi-agent orchestration and TinyTroupe (Microsoft's open-source persona-simulation library) to define realistic TinyPerson personas and place them in a simulated TinyWorld, plus a moderator agent representing the product to drive discussion. Transcripts were aggregated and analyzed for sentiment and themes (via NLP and LLM summarization). I refined persona realism with the client's marketing team using real customer profiles as guidance, and explicitly steered persona attributes away from demographic stereotypes— keeping everything in a secure environment with no sensitive real data.
The simulations produced fresh qualitative hypotheses—one persona consistently raised mobile-app security concerns, flagging a marketing emphasis the client had not prioritized. Because the method is synthetic, I labeled the output up front as hypothesis-generating, not evidence — its value was positioning these themes as concrete test items for the client's next real survey round, oriented their research budget, and demonstrated generative AI in a consultative role to their innovation lead.
The project sharpened my view of boundaries and ethics in AI simulations—careful prompt and agent design avoided stereotypes and inappropriate content, and the decision to present output as hypothesis material (never as validated insight) is the reason the client kept using it. Multi-agent simulations can spark intriguing results but must be validated with real data; they augment human creativity rather than replace user feedback.