Act Now or Wait: How Retail Agents Decide on Partial Evidence
I give talks and workshops on this material for retail teams and conferences. See past talks or invite me to speak.
Series: Foundations of Agentic AI for Retail, Chapter 3 of 14. Updated for the Third Edition (October 2026).
From the book: Foundations of Agentic AI for Retail, Third Edition
Chapter 2 left a store's pricing and inventory agents at 9:02 a.m., facing a low-stock product and a competitor's price cut. They could cut the price, reorder, or wait.
Before they choose, a promotion starts in the category, and sales spike in the East region, days before the next delivery. Is that real demand, cannibalization (one product's promotion taking sales from another), or a data error? Get it wrong and the store burns margin, runs out of stock, or both.
In Chapter 3, I give the agents what they need to make that call before the evidence is complete, and to know afterward whether it was right.
Why matching the competitor's cut can be the costly move
The instinct is to match the cut. In the chapter's example, that option drops out before any probability is needed. As the section on Bayesian decision theory puts it, "with stock this short, a cut gives away margin on units that would sell anyway."
That leaves rushing an order or waiting for the delivery. Bayesian decision theory treats the agents' view of the lift as a belief that moves with each piece of evidence, and weighs every action against it. The spike tips the agents toward ordering, and the chapter states how strong new evidence would have to be to tip them back.
That only works if the probabilities are honest. When I led AI at a bank in Canada, our churn and fraud models needed calibration, so that their probabilities matched how often things happened. When one lost it, the first sign was that it flagged too many frauds or too many churners. In a store, an overstated chance of real lift can trigger rush orders nobody needed.
Here is the shape of that morning's decision.
Why the best shelf space should not go to the priciest item
Some decisions come with a clear goal and clear rules. For those, the chapter uses optimization, which turns the goal and the rules into a model a solver can answer.
The shelf example makes it obvious. A short section has room for a blender, or for a kettle and a toaster side by side. A ticket-price rule hands the section to the blender. Rank by what each centimeter earns instead, and the cheaper pair wins by a wide margin.
The chapter's conclusion sums it up: "Optimization models set orders, prices, and shelf plans by contribution, not ticket price."
How to tell whether the agent's action actually worked
Two of the questions in the chapter's opening are causal: "Is demand dropping, or is it promo cannibalization?" and "Did the agent improve conversion, or did it only shift the mix?"
Both ask what would have happened if the agent had done nothing, which is the work of causal reasoning. The agent needs uplift, the part of the increase its action caused.
Correlation answers, "Did sales move?" Uplift answers, "Did sales move because of our action?"
— Foundations of Agentic AI for Retail, Chapter 3
Each tool corrects an instinct that feels right on a busy morning.
| The instinct | What the chapter shows |
|---|---|
| Match the competitor's cut | With stock this short, waiting beats matching whatever demand does |
| Give the best shelf to the priciest item | Ranked by what the space earns, a cheaper pair can win |
| A spike means real demand | Regional shifts and data errors cause spikes too |
| Sales rose, so the promotion worked | Only a comparison with doing nothing shows that |
What the chapter gives you
- The rush-or-wait decision on promotion morning, worked from the first spike to the evidence that would reverse it
- Five questions for choosing a decision framework, plus a guide that compares the methods of Chapters 3 to 5
- An experiment design template for agent rollouts, filled in for a pilot with holdout stores and stop rules
- Runnable companion code for a promotion test and a Bayesian product-recommendation agent, plus nine exercises
Go deeper
Chapter 3 of the Third Edition, out since October 2026, has the full method, the worked examples and the companion code. It also covers what this post only hinted at: the section on Bayesian networks, which link weather, promotions and demand, and the section on capstone metrics and causal questions, where three agents are judged by causal impact. New in the Third Edition: the forecasting tournament my team runs at OODARIS AI, which simple statistical models win for some product categories.