AI Assistant (Leat MCP)
July 20
→
Decision confidence
Every estimate includes a confidence range that shows how much evidence supports it. You decide how certain Leat AI must be before acting independently and what should happen when confidence is lower. Decisions can return to established rules, use a safer option, request approval, or proceed with less value at risk.
Propensity, uplift, and expected value estimates each include a measure of confidence. This reflects how much customer history is available, how recent it is, how many similar situations the model has seen, and whether the underlying signals point in the same direction. An estimate based on three transactions from a new customer is therefore presented differently from one supported by two years of regular visits.
Set the confidence required for Leat AI to act independently by incentive type, customer group, or location. Below that threshold, the decision can return to your existing rules, select the safest or lowest-cost permitted action, enter an approval workflow, or proceed with a lower-value incentive. Above the threshold, the system can act without waiting for review.
Confidence is naturally lower for new customers, recently launched incentives, new locations, and seasonal situations the model has not encountered before. Leat can defer these decisions or use a limited share of them to gather evidence within the boundaries you set. The number and cost of learning decisions remain capped, allowing confidence to increase through observed results rather than assumptions.
Each recorded decision includes the confidence behind it, allowing teams to see why the system acted independently or used a fallback. Model Monitoring compares stated confidence with actual results over time. If outcomes stop supporting the level of certainty reported, the model can be reviewed and adjusted before unreliable confidence affects more decisions.
Confidence threshold setting
Teams can set how certain Leat AI must be before acting independently for each incentive type, customer group, or location.
Fallback policy definition
Decisions can return to established rules, use a safer option, request human approval, or proceed with less value at risk.
New mechanic testing
New rewards and incentive types can begin with limited autonomy, then receive more as reliable results accumulate.
Learning limits
Teams can cap the share and cost of decisions used to gather evidence about a new customer group, location, or incentive.
Confidence accuracy & reviews
Stated confidence can be compared with actual results to identify where the model has become too certain or too cautious.
Deferred decisions
Teams can track how often decisions return to rules or human review and investigate increases for particular customer groups or incentive types.
