What a Day-Late Reorder Teaches You About Retail AI Agents
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 1 of 14. Updated for the Third Edition (October 2026).
From the book: Foundations of Agentic AI for Retail, Third Edition
In Chapter 1, I hand one low-stock product to an agent small enough to read in full. Each simulated day it reads the stock, compares it with a threshold, and orders or waits.
On day three, the stock sits exactly at the threshold. The rule reorders only below it, so the agent waits, and that day the product sells three times its usual amount.
On day four it sees the gap and orders. Nothing in the code is broken. The order still came a day late.
That small miss carries the chapter's argument. An agent turns store signals into actions through a loop, and a retailer should want one where delay is costly and every action is tied to policy, evidence, and an escalation path.
What separates an agent from the automation you already run
Strictly speaking, that little agent is automation, an agent only in the classic sense. Agentic AI adds autonomy to the rule-based and predictive systems retailers already run, deciding what needs doing next without waiting for a prompt at each step. Unlike generative AI, which produces content on request, an agent is goal-directed: it plans, calls tools and keeps state to carry multi-step work through to an action.
The book's teaching model is the Perceive–Reason–Act–Learn loop, and the demo's rule falls short of it at every step.
| Loop step | The demo's reorder rule | An agent in the book's sense |
|---|---|---|
| Perceive | One number, its own stock level | Many streams, from sales to supplier lead times |
| Reason | Compares stock with a fixed threshold | Forecasts demand instead of trusting a fixed threshold |
| Act | Orders with no policy check and no trace | Acts under policies, budgets and monitoring |
| Learn | Nothing, the step is left empty | Evaluates outcomes and sends gated updates back |
It ordered after the spike instead of before it, because a rule fixed at 50 cannot ask whether 50 is still right.
— Foundations of Agentic AI for Retail, Chapter 1
The two silent killers hiding in a demo that works
The demo runs cleanly, which is why it makes a poor model for production.
In my experience, two silent killers show up early in agent projects. One is automation without a control plane, where actions ship faster than anyone can check or audit them. The other is a feedback blind spot: an agent ships with no real evaluation loop, and slow drift goes unnoticed.
The reorder agent has both, by design, in two places on its loop.
The book's answer is the Retail Agent Operating Model, or RAOM, which pairs the loop with a control plane. Before a real order goes out, RAOM says which policy permits it and where it escalates if something goes wrong. Chapter 1 introduces it, and the rest of the book builds it.
How to tell whether a decision deserves an agent
On the first day of class, I tell my students that agentic AI is the most important topic they will learn. A retailer needs a narrower question. In "Making the Retail Business Case," I put it this way: "Which decisions justify autonomy, and what evidence would make the investment worthwhile?"
Where the steps are fixed, a workflow is the better tool. An agent earns its place where picking the next step from changing evidence adds value you can measure.
Measuring it is the hard part, and no survey can establish the value of one deployment. So the chapter closes the business case with an investment test, four things to agree on before you fund a pilot. One is a stop rule, the result at which the pilot ends.
What the chapter gives you
- The Perceive–Reason–Act–Learn loop, the teaching model the rest of the book runs on
- A one-product inventory agent with companion code and a seven-day run you can trace
- A first look at the Retail Agent Operating Model (RAOM), plus a layered reference model for finding a pilot's gaps
- The investment test, and ten exercises from quick checks to architecture design
Go deeper
Chapter 1 of the Third Edition, out since October 2026, has the full loop, the agent's code and its day-by-day run, and the complete business case. It also covers what this post only hinted at, the layered architecture in "Agentic AI System Architecture." New in the Third Edition: Part I follows one low-stock product from this first agent to the pricing and inventory agents that share it.