Agentic AI in Retail: Adeo and Decathlon's Playbook

Key takeaways

1. The framing sentence. A human reading a poor definition compensates with experience; the agent applies it as given, including when it holds for France but not for Spain and nobody wrote that down.
2. Platformization was not built for agentic AI. It targeted Leroy Merlin's marketplace (60,000 in-store references → several million online) and Decathlon's expansion (a few dozen → more than 50 countries). Its byproduct — a cart API, a customer API, a payment API, with the right contract — is exactly what an agentic model expects. Taillade adds standardization: bringing everyone onto a common platform cleaned up fifty years of accumulated layers.
3. Two revolutions in parallel. A company that has not finished its transition from commerce-IT to platform-IT must run the cultural and the technological fronts together while AI compresses the timeline. A direct echo of the amplification thesis in [[dora-google-cloud-roi-ai-assisted-software-development-j-curve-2026-04-21]].
4. Agentic commerce, effect 1: the retailer no longer builds the journey. The customer's agent decides the order of the discussion, can start at the end, return to the middle, negotiate — the foundation must absorb this type of interaction, including at the level of calls per second.
5. Effect 2: the question Taillade is tracking. "How long before we could disappear from this conversation?" "A Kiprun t-shirt for my marathon" keeps Decathlon in the loop; dropping "Kiprun" opens the field to every running retailer, and the customer will always make the least effort.
6. Grymonprez's counterpoint. the same fear at the launch of Google Shopping, "we've never done so much e-commerce." The customer delegates the simple and low-risk, not the expensive bike or the hardwood flooring; half of spending at Leroy Merlin comes from projects that take weeks to plan. Hence the rule: "if tomorrow it's easier to get answers about home improvement from ChatGPT or YouTube than from me, the problem is mine."
7. Internal reconfiguration, not layoffs. API/event workflows are rigid, an agent reasons about the next step; some roles will be compressed, accelerated, or shifted. Power shifts to problem solvers — "innovation comes from someone who has a problem" — with IT keeping the big issues (supply chain, millions of references), provided everyone else produces within the group's guardrails.
8. Workstream 1, the harness. development environment, sandbox, skills. The urgency comes from how easy access has become — "you get a Claude license, you can already get started […] but then you won't be doing it in the right harness." Instruction to the teams: "let's provide a framework, even if it isn't perfect."
9. Workstream 2, zone-based security. zero agentic code on payment platforms, the checkout flow, and supply chain movements until proven otherwise; a full green light wherever risk is nil.
10. Workstream 3, data semantics. (a debt Adeo acknowledges): a single definition, understandable by an agent, for every business concept, tackled by vertical — finance then supply — rather than laterally. Taillade: human knowledge used to compensate for missing documentation; organizing knowledge becomes an activity in its own right.
11. The precedent of 150,000 dashboards. Adeo opened up all its datasets to self-service BI; after a year or two, more than 150,000 dashboards and nobody speaking the same language. A return to certified data products and data contracts, for lack of a framework set from the start.
12. The fragmentation scenario. An agent unaware of the group's payment capability wires up Stripe, then another one does too, and by the thirtieth integration the IT system is fragmented: "your IT system can collapse under its own weight." Coding agents pick up the ingredients of low-code with far greater production capacity — hence the shift of responsibility toward IT: tooling and a framework "that leaves no room for shortcuts" replace training and peer vigilance.