AI Assistant (Leat MCP)
July 20
→
Decision simulation
Try a new strategy on last quarter before you try it on this one. Change the objective, the threshold, or the options available, and see how recent decisions would have come out differently. Because the simulation is a replay of real decisions and real customers, what you see is what would actually have happened, not a guess.
A different objective, or different weights between objectives. A higher or lower confidence threshold. A new action added to what the engine may choose, or an existing one removed. A more or less aggressive suppression setting. A different split of budget across mechanics. Anything that changes how decisions are made can be simulated, which is the part of the program that rule simulation doesn't reach.
Leat takes a period of real decisions, with the real customers, contexts, and options that existed at the time, and re-runs them under the proposed strategy. For each, it shows whether the outcome would have differed: the customer who'd have received a message instead of a voucher, the one who'd have received nothing, the one who'd have received more. Projected results come from the same models the engine uses, with the same confidence ranges, and where a holdout ran in the period, the projections are anchored to what actually happened.
A shift toward margin lowers redemptions by this much and raises margin by that. A tighter confidence threshold defers this many more decisions to rules and avoids this much estimated waste. Adding a new mechanic changes the mix like so, by segment. Simulation shows what a strategy gives and what it costs, per segment and per location, so the decision about the decisioning is made on a trade-off you can see rather than a preference.
Simulation is an estimate. A strategy that looks better in simulation is a candidate, and the way to know is to run it as an A/B test against the current one on a live split. What tests well rolls out, and the simulation that predicted it is checked against the result, which is how simulation itself gets more accurate. Every scenario, its projection, and what happened when it went live is recorded, so the program's strategy has a history and not just a current setting.
Program change management
Change a rule once and have every campaign, channel, and decision honor it
Responsible AI governance
Set the boundaries the system operates within and prove it stayed inside them
Promotion governance
Hold eligibility, spend, and frequency limits in one place and apply them before any incentive is issued
Margin-safe promotions
Run offers that can't combine into a loss
Location budget management
Give every site or franchisee an allowance that operates inside the group cap
