Why a Retail Agent Pilot That Passes Still Needs a Post-Mortem
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 13 of 14. Updated for the Third Edition (October 2026).
From the book: Foundations of Agentic AI for Retail, Third Edition
A regional retailer is heading into a holiday weekend with tight inventory and rising promotion pressure. A new agent is helping its replenishment team, and it has flagged one product, SKU-123: 20 units on hand against 60 units of expected weekend demand.
The agent proposes a supplier order. Policy will not let it commit that much money on its own, so the order waits for a buyer, who approves it. The weekend comes, and every unit of demand is met.
The retailer is hypothetical and its record is synthetic, but its question runs through every playbook in the chapter: when may this agent be trusted with more? Chapter 13 answers it in five retail domains, with the same answer in each: keep the agent bounded, and widen its authority only as far as the evidence allows.
What a passing record still leaves open
I follow SKU-123 through the whole chapter, and its record does not stop when the evaluation passes.
The section "Worked Record: SKU-123 Before the Holiday Weekend" shows why. The agent ordered 80 units to cover a 40-unit shortfall, so 40 units were still in stock when the weekend ended. Nothing in the record says whether post-holiday demand, a case pack or a minimum order quantity justifies the larger order.
The evaluation passes, yet the record shows the trade-off a buyer would question first.
— Foundations of Agentic AI for Retail, Chapter 13
The chapter closes with SKU-123's post-mortem, which shows what one synthetic weekend can and cannot tell the buyer.
| What the record shows | What it leaves open |
|---|---|
| A buyer other than the requesting planner approved the order | Whether an approval stays bound to what the buyer actually saw |
| A supplier call that timed out was retried safely | Recovery when it is unclear whether the supplier accepted |
| Every unit of weekend demand was met | Whether the leftover stock pays, and value across representative weeks |
What actually breaks when an inventory agent meets real data
SKU-123's 20 units are a record, and records drift from the shelf. The section "Benefits, Challenges, and Lessons Learned" says it plainly: "The challenges lie mostly in operations and data rather than in models."
I have lived that one. In my team's planning system at OODARIS AI, beginning-of-period inventory could be negative, and returns could push stock below zero, leaving the agents to work out how to attribute them. Each tenant also ran its own calendar, which the agents had to understand and adapt to. Both are questions about what the data means.
How one pattern carries from the stockroom to the checkout
From inventory, the same holiday weekend moves through four more domains.
A pricing agent has to protect margin without raising stockout risk on the SKUs already flagged. A paid placement could send shoppers to a product the replenishment agent knows is short. Shoppers may ask the store's assistant about those same SKUs. In agentic commerce, a mistake reaches a shopper's payment, and only a cancellation or a refund undoes it.
Two of the five playbooks, retail media networks and agentic commerce, have no case studies behind them, so the chapter presents them as design sketches. All five share the same parts and map to the Retail Agent Operating Model (RAOM).
Each playbook starts bounded, and each pilot ends in the same review.
My favorite line, from the "Customer-Facing Agents" section, belongs on every service dashboard: "a deflected contact counts only if the problem is solved."
What the chapter gives you
- Five playbooks, from inventory to agentic commerce, built the same way and mapped to RAOM
- The SKU-123 record, a synthetic replenishment episode you follow from the agent's proposal to its post-mortem
- A case snapshot of published retail accounts and anonymized field composites, each read for what it actually establishes
- Runnable companion code for three demos (an inventory agent, a pricing agent and a shopping assistant), and eight exercises, including a decision record you draft yourself
Go deeper
Chapter 13 of the Third Edition, out since October 2026, has the full playbooks, the SKU-123 record with its post-mortem, and the companion code for each demo. I only hinted here at the pricing section's evaluation card, written before a pilot starts, and at who owns each step in the agentic commerce playbook when an accepted order has to be unwound. New in the Third Edition: Part V's hypothetical holiday pilot, and what it must show to earn more autonomy.