Who Answers When Your Pricing Agent's Markdown Looks Unfair
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 12 of 14. Updated for the Third Edition (October 2026).
From the book: Foundations of Agentic AI for Retail, Third Edition
A pricing agent marks down one style of wool winter coat by 40% in one region's stores, to clear stock that is selling too slowly. The markdown crosses a policy threshold, so it waits for the outerwear category merchant, who approves it within the deadline.
Shoppers in the next region see the lower price in the retailer's app. The next morning, customer service faces a wave of complaints about unfair offers.
The agent did what its policy allowed and the merchant signed off, yet the head of pricing, who owns the agent, now has to answer the complaints.
The case is hypothetical, and it carries the chapter's big idea: governance is how a brand prepares its answer before the agent acts.
Why the agent can't be the one who answers
The agent followed its rules, yet the brand still owns the outcome.
— Foundations of Agentic AI for Retail, Chapter 12
An agent can act on its own, but legal and ethical responsibility stays with the people who deploy it. So the chapter gives every decision a named owner, with a deputy and a deadline, from the head of pricing who owns the policy to the customer-care lead answering complaints.
At a bank in Canada where I led AI, an accounting firm audited and reviewed our models for three months. I could answer every question, because the dashboards, reports, audit logs and explanations were already in place.
The chapter treats governance as a control plane that checks permission before the agent acts and keeps the evidence afterward, shaped like this for the coats.
What the approval card showed, and what no rule checked
A discount of 40% needed explicit approval, so the proposal reached the merchant as an approval card with its evidence and 24 hours to decide.
Read the card again after the complaints, and two gaps show. The policy checked the 40% change, but a loyalty coupon stacks at the till, so members would pay 46% below the price before the markdown. And the card says nothing about the region next door.
Oversight has failure modes too. Once people trust the agent, an approval queue tends to become a rubber stamp. My team at OODARIS AI treats that as a sign of where to look, but the chapter is careful: how often people approve is no evidence, on its own, that review can safely be removed.
Why an approved markdown can still be the wrong call
In the section "Capstone Governance Flows", the head of pricing separates the questions the complaints had run together. The audit trail settles some of them and leaves others open.
| Question | What the record shows |
|---|---|
| Authorization | Settled: it went through the approval policy, and the merchant approved it in time |
| Rationale | Recorded stock cover and season explain a clearance; reproducing the exact proposal needs a replay |
| Commercial call | Open: on the chapter's forecasts, another option recovers more |
| Fairness | Open: the fairness audit hasn't run, and the coupon was never checked against any limit |
That other option, moving coats in the sizes the next region was short of, would on the chapter's forecasts have recovered more than the markdown and kept both regions at full price. The agent couldn't propose it, because it had no tool for a transfer.
So the head of pricing makes a provisional decision, pausing the agent's markdowns on that style until set conditions hold. The chapter then ends the case with go or no-go governance gates: if any gate is red, the launch waits.
What the chapter gives you
- The winter-coat markdown, followed from the regional dossier and the merchant's approval card to the head of pricing's provisional decision
- A map of who answers for each part of an agent's decision, with levels of autonomy and escalation protocols for what rules can't settle
- Go or no-go governance gates for launch, and a risk section on fail-safes and circuit breakers
- Companion code for a small policy engine and a human-in-the-loop approval queue, plus exercises that end in a one-page memo on the coats
Go deeper
Chapter 12 of the Third Edition, out since October 2026, has the full method, the coat case with its numbers, and the companion code. It also covers "Transparency and Explainability," on what a regional manager needs to see behind a markdown, and "Risk Management for Autonomous Systems," on fail-safes for the moments no reviewer is watching. New in the Third Edition: the winter-coat markdown as the capstone's governance test, and how an approval queue becomes a rubber stamp.