- About 77AI calls for the whole population. A naive build needs a million
- 20 roundsOf reaction and mutual influence, every run
- 2.25×Loss aversion, per Prospect Theory, over every agent
- 6Verticals with their own calibrated playbook, six more mapped
The challenge.
Most businesses analyse a decision after they have made it.
You raise prices 15%. You launch the product, or cut the service. Then you wait, and the market tells you what it thinks, and by then the decision is already spent.
The tools built for that moment are thin. A spreadsheet models the arithmetic of a decision but not the feeling of it, and the feeling is what churns customers. A survey asks people what they believe they would do. A focus group asks a handful of them in a room. None of them show the part that actually hurts: what happens when the people who are angry reach the people who are undecided, and a survivable amount of unhappiness turns into a wave.
The reason nobody simulates it properly is arithmetic. A genuinely diverse population of half a million customers, each one reasoning about your specific decision, needs on the order of a million AI calls. That is around $360 of compute a run, slow enough that nobody would run it twice, which kills the idea before it starts.
If that shape looks familiar, the consultation is where we go and find yours.
Start with the free consultationWhat we built.
Foresay rehearses the decision instead.
You describe it in plain English in a short wizard. Foresay builds a population of up to 500,000 simulated customers, grounded in real demographic and economic data, and runs them through roughly twenty rounds of reacting to the decision and influencing each other.
What comes back is a read you can argue with. The risk. The churn. The sentiment. What people actually say. Where the ones who leave go. The single change that most reduces the backlash. And a verdict on how much the engine trusts its own answer.
Five stages get it there. The decision is parsed and classified by type, magnitude and vertical. A population is generated from real demographics and the current economy. Five to eight AI calls capture how each customer archetype reasons about this decision in this economy. The whole population then reacts and influences its neighbours across the rounds, and a calibrated read out returns the result.
The third stage is the engineering result. AI spend is driven by the number of archetypes, not the number of agents, so reasoning captured once propagates across all 500,000 at no further AI cost. Half a million agents run on about 77 calls. The behavioural maths still runs over every single one of them and is never sampled, which is why the cost per agent falls roughly thirteen times going from ten thousand agents to five hundred thousand.
The population is built rather than sampled. Ages come from Census brackets and income from Census ACS distributions, with the archetype mix from consumer behaviour research, so at half a million it reproduces actual demographic skew. Macro data, including unemployment, inflation and sentiment from FRED, plus UK ONS and Bank of England figures, is pulled per location and baked into every agent's baseline, so changing the economy changes the prediction. And individuality comes from the maths: each agent's own behavioural scoring, its position in the social network, and per agent noise, so no two react alike inside one archetype.
Two published models then run over the whole population, deliberately in the open. Prospect Theory gives the decision maths, with loss aversion at about 2.25×, which means a loss lands about 2.25 times as hard as the same size gain. That is why a 15% rise stings far more than a 15% discount delights, and why a spreadsheet cannot feel it. DeGroot opinion dynamics gives the social maths: every round, each agent's view moves toward the people who influence it, so consensus, polarisation and viral backlash emerge from the network instead of being assumed. The archetypes are few and named. Loyalist, value hunter, sceptic, early adopter, social follower.
Naming all of that is safe, because the ingredients are not the engine. Anyone can reach for Prospect Theory, opinion dynamics and census data. What cannot be read from outside is the calibration that makes them behave like a market rather than noise, and it took roughly 1,800 hours of computation and structured testing to find, graded against a benchmark that scores every version of the engine against reality.
The results.
The most interesting thing about it is what it does when it is not sure.
It abstains. On the calls it does make, error roughly halves. The promise is deliberately not one blanket accuracy number. It is high accuracy on the confident calls, plus honesty about the rest.
On severity, meaning how big the reaction will be, it lands within one tier 85 to 91% of the time in internal benchmarks, against roughly 25% for chance. On direction, meaning welcome or backlash, it passes 90% in internal benchmarks, and treats direction as the easier axis and as supporting evidence rather than the headline. Those figures are internal benchmarks on held out and hard boundary cases. Nothing there is externally certified, and the engine does not present it as if it were.
It publishes its own ceiling too. Prediction from the wording of a decision alone tops out near 0.70 correlation with real outcomes, because execution, timing, competitor moves and luck are not in the wording. That ceiling was established several independent ways, including a study of 270 company outcomes.
Six verticals carry their own calibrated playbook, with confidence reported per vertical rather than averaged into one flattering number: retail and small business, banking, insurance, nutrition and health, venture capital, and forecasting. The architecture maps to six more.
Depth is a dial. Every run blends AI reasoning with behavioural maths, and both the AI share and the model tier are levers, so a deeper blend buys a richer read. A REST API runs simulations programmatically, so the engine can sit inside another product entirely.
The name is the idea. Fore·say, like foresight. To hear the reaction before it happens.
Tell us what yours looks like.
Bring us the part of your business that runs on someone remembering. We will tell you what it would take to build it out.