Recent research shows LLMs can simulate human responses in social experiments with a high correlation to real effects (r ≈ 0.85 across an archive of 70 studies with representative samples). In financial services, that unlocks a new paradigm: virtual markets of millions of synthetic customers to forecast product uptake, adoption risk and regulatory impact.

3D diagram: public data, persona generator, synthetic agents, simulator, ecosystem and analytics.
Figure 1 — Conceptual architecture: from public statistics to a synthetic population and large-scale simulations.

Why it matters

Banks and fintechs still rely on surveys, focus groups, A/B tests and classical statistics — costly approaches with limited samples and weak coverage of novel scenarios. Synthetic Agents offer a complementary path: a digital population that stands in for millions of consumers and enables cheap, fast experimentation.

Profile of each synthetic agent

DimensionAttributes
DemographicAge, region, education, occupation
FinancialIncome, credit score, indebtedness, repayment capacity
BehaviouralRisk aversion, price sensitivity, credit propensity
PsychologicalConservative, moderate, aggressive, planner

Banking agent examples

Retiree (social security)

Age 68, benefit income, low digital affinity. Simulations: payroll-linked credit, biometrics acceptance, rate sensitivity, reaction to new rules.

Digital native

Age 25, app-first, limited net worth. Simulations: digital account, instalment payments, cards, automated investing.

SME and high net worth

Variable cash flow (working capital, receivables) or high wealth (wealth products, funds, structured credit).

Use cases

  • Credit — “What if we cut the payroll-loan rate by 0.5%?” → uptake, volume, P&L impact
  • Launch — expected conversion, objections, age bands and best channels
  • Messaging — compare copy (“Rates from 1.5%” vs. “Save up to $50 a month”)
  • Payroll credit — campaign design, regulatory shocks, digital signature friction by age

The bank as an agent ecosystem

Beyond customer agents, the model expands to specialist agents: regulator, competitor, market (macro), fraud and banking correspondent — interacting in virtual societies where opinions spread and adoption emerges organically.

Thesis: Synthetic Agents can be for banks what flight simulators are for aviation — a safe environment to rehearse decisions before they touch real customers. The paradigm shifts from “launch and measure” to simulate, learn, then launch.

Scientific foundation

The approach aligns with research on Generative Agents, synthetic populations and LLM-based forecasts of experimental effects. Limits remain: demographic bias, out-of-distribution contexts and misuse risk — so governance and validation against real data stay mandatory before critical decisions.

Primary reference: Hewitt, L.; Ashokkumar, A.; Ghezae, I.; Willer, R. Predicting Results of Social Science Experiments Using Large Language Models. Stanford University & New York University, 8 August 2024. Evaluation archive: 70 pre-registered nationally representative experiments (TESS programme and replications), 476 treatment effects and 105,165 participants; correlation between GPT-4 simulated and actual effects of approximately r = 0.85.
← All publications