One Gate at a Time: How a Retail Agent Pilot Earns Autonomy
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 14 of 14. Updated for the Third Edition (October 2026).
From the book: Foundations of Agentic AI for Retail, Third Edition
The head of replenishment has a gate to sign. The evidence is thin: one SKU, one simulated holiday weekend, one large supplier order that policy held for the buyer, and one approval.
The agent did what it was asked, and every unit of weekend demand was met. The review still says hold.
Chapter 14 opens with that signature because it raises the book's last big question: how does a retailer get from a first pilot to dependable autonomy, and in what order? My answer is to hand decisions to software one gate at a time, as far as the evidence carries you.
Why a hold is the gate doing its job
The chapter places the pilot on a maturity model with four stages, and every move up has entry prerequisites and exit gates. The section "Implementation Roadmap and Maturity Model" is strict about what a gate is: "Each exit gate is an acceptance test, not an installed control or a documented metric."
Whatever the stage, a gate review has the same shape.
A hold is the gate doing its job.
— Foundations of Agentic AI for Retail, Chapter 14
What happens when the agent wins on value and still fails
The second gate review, a seeded simulation built on the companion code (synthetic, not field evidence), runs the agent against the incumbent ordering rule.
The agent beats the rule on value, and it ties up less capital and writes off fewer units. It also runs out first on some holiday weekends when demand runs well above forecast, and it adds time to the buyer's review. Two gates fail, so the verdict stays hold.
The tempting move is to relax the availability bar, since the miss is small. The section "The Second Gate Review" closes that door: the owner set the gate before measuring, and lowering it "after seeing the value would fit the threshold to the answer."
Other shortcuts look just as reasonable in the moment.
| Tempting shortcut | What the chapter asks for |
|---|---|
| Treat the buyer's approvals as proof of maturity | Approval alone does not settle the stage |
| Compare the pilot with an industry benchmark | Judge it against its own stores, measured the same way |
| Read a simulated pass as field proof | Even a full pass needs field evidence and the owner's signature |
Why the people around the agent set the pace
One of the chapter's eight pitfalls holds a prediction I got wrong. I thought people would appreciate the productivity gain and business value that agents bring far more than they do. I expected some resistance, but not resistance so general.
None of my team's rollouts at OODARIS AI has failed, yet some users do not know how to prompt and converse with the agents, and they can still make their changes by hand. The chapter's answer is to ask what each objection is about and meet it with evidence.
What will still hold when the tools have changed
Today's tools will age, and the chapter treats emerging technologies as hypotheses that must earn adoption with evidence. In "Conclusion: Charting the Course for Agentic Systems," I name what I expect to outlast them: "the mathematics of decisions, what counts as evidence, who has the authority to act, how these systems fail, and how to tell whether a change worked."
This is the last chapter, and the last post in this series, so thank you for reading along. Chapter 1 asked when to hand a decision to software, and this chapter finishes the answer. At any gate, as the chapter's conclusion puts it, the answer can be "not yet."
What the chapter gives you
- A maturity model for agentic retail, with entry prerequisites and exit gates for every move up
- Two gate reviews of the running pilot, the second a simulation you can run from the companion code
- Five success factors, eight common pitfalls with ways to avoid them, and a build, buy or partner framework
- A watchlist of emerging technologies and ten exercises, including a next gate review you write yourself
Go deeper
Chapter 14 of the Third Edition, out since October 2026, has the full maturity model, both gate reviews with their evidence, and the companion code behind the second. I only touched on the build, buy or partner framework and on the horizons section, where each milestone depends on conditions more than on dates. New in the Third Edition: a prediction I got wrong, about how much people would welcome what agents bring.