In distribution, a promotion is usually closed on one number: sales during the campaign window. That number is real, but it is not effect. Effect is how much of that volume would have happened anyway, and how much of it smothered a neighbouring SKU on the same shelf. Without that split, budget follows the campaign that looks busy, not the one that created net sales. This is exactly the split agentic AI performs on real FMCG distribution invoices, not a presentation slide.
Baseline before mechanic
A promotion mechanic (tiered discount, bundle, gift-with-purchase) is only comparable against a shared baseline: same brand, same channel, same seasonal window. Comparing Yalda with a summer festival without that alignment is a calendar contest, not a test. Didrah’s intelligent agent builds effect from invoices and campaign lines, not from marketing’s closing slide.
Measured uplift is not revenue
Measured uplift can be zero or negative while window sales are still up. That pattern is usually cannibalization or pulled-forward demand. If a campaign is fifty-six percent overspent with no measurable sales effect, the right call is not to rerun the mechanic; it is to stop and redesign. This is where the autonomy ladder matters: the agent can surface the finding, but resetting the budget is still a human decision.
- Before raising budget, ask for effect by mechanic, not by campaign slogan.
- Do not rank an unmeasured campaign against a measured one; finish measurement first.
- If effect is near zero, the next question is cannibalization, not brand awareness.