AI Assistant (Leat MCP)
July 20
→
Model monitoring
Catch a model drifting before your customers do. Compare what was predicted against what happened, continuously, for accuracy and for fairness. Because drift automatically widens confidence and defers more decisions to your rules, a model that's losing its edge loses autonomy before it loses your trust.
When the engine said a customer had a 70% chance of returning, did 70% of customers like them return? When it estimated an offer would lift purchases by twenty points, did the holdout show twenty? Calibration is checked continuously, per model, per segment, per mechanic, using the outcomes and holdouts Optimization already produces. Accuracy is a chart that updates, and where the line bends is visible the week it bends.
Two kinds of change are watched. The inputs shifting, so customers now behave differently from the ones the model learned from. The outputs slipping, so predictions that used to hold no longer do. Either crosses a threshold you set and raises an alert to the people you name, with the model, the segment, and the size of the gap. A drift found at 3% is a recalibration; found at 30% it's a quarter's budget.
The fairness definitions set under Responsible AI controls are checked against outcomes here. Offer rates, average incentive value, and access to top rewards are compared across the groups you specified, and a gap widening toward your threshold is flagged before it becomes a pattern. A model can drift toward unfairness as easily as toward inaccuracy, and both are watched the same way.
Drift isn't only reported; it changes behavior. When a model's stated confidence outruns its accuracy, its confidence intervals widen automatically, so more decisions fall below the threshold and defer to your rules until the model is recalibrated. Adaptive decisioning retrains on the recent outcomes. A model can be paused or rolled back to a prior version. Every finding, alert, and response is recorded, so when a compliance team asks how the system is kept honest, the answer is a log.
Responsible AI governance
Set the boundaries the system operates within and prove it stayed inside them
Program integrity
Keep stamps, points, and rewards worth what they're meant to be worth
Program change management
Change a rule once and have every campaign, channel, and decision honor it
Margin-safe promotions
Run offers that can't combine into a loss
Data subject requests
Handle access and erasure requests in one action across the whole platform
