Making decisions with synthetic populations: the practical guide.

Organizations prepare their decisions far from the people concerned: on averages, dashboards and internal convictions. Yet it is in the encounter with humans, customers facing a new price, citizens facing a reform, teams facing a transformation, that the decision succeeds or fails. A synthetic population moves that encounter upstream. The decision-maker can interview the people who will live the impact of the decision, compare decision A against decision B, and observe the trajectories each one produces over several months or years.

The change is not only a matter of speed. It is a change in the nature of the analysis. A classic study returns an average snapshot at a point in time. A simulation at scale returns the fine internal structure of the population: the stable groups, the hesitant groups, the weak signals, the threshold effects, the contradictions between neighboring groups. A project that looks moderately acceptable on average can be rejected by a precise group, on a precise territory, for a precise reason, with a measurable blocking risk. In one completed public policy case, 4% of the population concentrated 78% of the political risk of the measure: no average could have shown it.

The criterion is twofold: strong behavioral uncertainty and a high cost of error. The decisions that lend themselves best are those where the reaction of populations conditions the outcome: setting or changing a price, launching an offering or a brand, repositioning a product, transforming an organization, reforming a public program, responding to a crisis, anticipating a market trajectory, arbitrating between deployment architectures. The 45 completed cases published on this site cover eight decision types, from launch to crisis response, across twelve industries.

Conversely, decisions whose outcome does not depend on human behavior, a purely technical optimization, an infrastructure choice with no adoption dimension, belong to other tools. The synthetic population illuminates the human part of the decision: that is its purpose and the limit of its competence.

A simulation is worth what its question is worth. The useful formulation is not a theme (“our pricing policy”) but a dated decision with its options (“announce an 8% increase in January, all at once or in two steps, with or without a support mechanism for vulnerable customers”). From that formulation follow the relevant populations: who lives the impact, on what territory, over what horizon. This framing determines the depth of everything that follows. On our system, the dynamic agents conduct this framing with the decision-maker, from the initial brief to the first deliverable, in 20 to 30 minutes.

The most common misuse is submitting an already-made decision for validation. The value comes from comparison: a useful simulation confronts 4 to 9 genuinely contrasted scenarios, varying the structuring parameters, the amount, the calendar, the announcement sequence, the support mechanisms, the messengers. The completed cases show it again and again: the dominant strategy is rarely the initially favored option, and often a combination no one had formulated. Testing wide costs a few extra minutes of compute; deciding narrow costs months of correction.

The output takes the form of a queryable investigation file, structured around four objects. The insights: the dynamics identified in the interviews, often counterintuitive, always attached to the typologies that carry them. The compared scenarios: the projected trajectories of each option, with their measured gaps. The tipping points: the thresholds, the critical windows, the moments where a trajectory changes regime. The reasoning paths: the complete trace that connects each conclusion to the individual interviews that produced it.

Two reading reflexes make the difference. First, go below the average: the granular typologies, 10 to 30 per simulation, are the granularity at which the decision plays out; an average adoption of 61% can conceal an absolute rejection within the minority that holds the blocking power. Second, contest: every insight must be traceable back to the interviews, questioned, corrected. An investigation file is not a closed answer, it is an instrument of exploration; it lets itself be audited by risk and compliance departments.

The simulation complements the other sources, it does not replace them. When observed behavioral data exists and is relevant, it remains the reference: the synthetic population deploys its value where real data does not yet exist, a new product, an untouched territory, an unprecedented configuration, or does not suffice, behaviors under unobserved scenarios. Likewise, internal expertise is not disqualified: it is tested. The teams’ convictions become hypotheses submitted to the population, and the disagreements between internal intuition and simulation are precisely the places to dig.

The decision remains human. The investigation file illuminates the arbitration, it does not make it. Once the decision is made, the simulation keeps serving: the projected trajectories become a vigilance baseline, against which the observed indicators are compared week after week. The completed cases document gaps between projected and observed trajectories below 8% on average over the first months; when the gap widens, it is a signal for adjustment, not a failure of method. The most advanced organizations industrialize the cycle: simulate, decide, deploy, measure, recalibrate.

The first misuse is mechanical prediction: treating the projections as certainties. They are probabilistic; they reduce uncertainty without eliminating it, and a major exogenous event can invalidate a trajectory. The second is substitution: discarding available real data in favor of the simulation. The third is delegation: letting the investigation file decide. An organization that delegated its decisions to a simulation system, whatever it may be, would commit a major methodological and political fault. The tools augment the decision-maker’s lucidity; they do not replace their responsibility.