AI Assistant (Leat MCP)
July 20
→
Uplift modeling
Predict the difference between what is likely to happen if you act and what is likely to happen if you do not. Leat uses results from your own controlled tests to estimate which customers an action could influence, where it would make no difference, and where it could make the outcome worse.
Uplift modeling produces two estimates for each customer and candidate action: the likely outcome if the action is taken and the likely outcome without it. This separates customers who would act either way, those unlikely to act either way, those who may respond negatively, and those whose behavior the action could genuinely change. The final group is where an incentive is most likely to create additional value.
Uplift is learned by comparing similar customer groups, where one group receives an action and another does not. Leat uses the holdouts and controlled tests run through Optimization, allowing estimates to reflect the behavior of your customers and the actions in your program. Each completed test provides more evidence for future estimates.
Next Best Action uses uplift to estimate what each candidate could change. If every available action is unlikely to influence the customer enough to justify its cost, taking no action can rank first. Uplift can also show whether a €10 voucher is expected to achieve more than a €5 voucher or whether the extra value would be unnecessary.
Each uplift estimate is a prediction that can later be tested against results from comparable holdout groups. Optimization measures the additional outcome created by the action, then compares it with the expected uplift. Models can be adjusted when estimates are consistently too high or too low, while Model Monitoring tracks whether their accuracy changes over time.
Uplift test design
Holdouts and controlled tests provide treated and untreated customer groups from which the model can learn.
Incentive value calibration
Teams can compare incentive amounts and use the lower value when it is expected to change behavior just as effectively as a higher one.
Do-not-disturb identification
Customers with little or negative expected uplift can be excluded from actions that would waste value or reduce the chance of the desired outcome.
Model recalibration
Predicted uplift can be compared with measured results so models can be adjusted when estimates become too high or too low.
See uplift by segment or group
Teams can see where incentives change behavior and where they are unnecessary across customer groups and program activities.
Action comparisons
Rewards, vouchers, messages, and other actions can be compared according to the additional outcome each is expected to create.
