Agentic AI in Retail: Adeo's CDO on Lock-In & FinOps

Key takeaways

1. Make vs Buy. at Adeo is « at least 70% » make: no competitive differentiation in full buy.
2. Every transformation is won on two fronts at the same time: culture AND structure. The same playbook (which drove digital transformation: waterfall → agile, product, platforms, tech radar, APIs, microservices) is being replayed with AI.
3. Accountability. remains human and tied to organizations, all the way to legal: « I don't want some guy telling me it's broken in production but it's not my fault, it's the agent ».
4. Documentation debt. (mediocre product docs, nonexistent process docs) « is going to slap us in the face very soon »: an agent can't read documentation that doesn't exist.
5. Mathieu never talks technology at the board: he talks customer experience, ROI, profitability, investment.
6. He doesn't even ask for extra budget for AI: the gains (compressing JIRA tickets, ops) fund the new work (« at flat »). A reuse logic: the time saved is reinvested in service of the in-store salesperson / the customer.
7. No fantasy of « firing all the developers thanks to Claude Code »: on the contrary, he anticipates an avalanche of requests — P10/P15 projects move to P2 because « no project costs too much anymore ».
8. His big « wake-up call »: Claude Code (recreating a VM without having coded in a long time), then seeing Karpathy and others « ship » more, then his personal agent « Hermes » (with memory).
9. Token cost. confident it will fall as cloud FinOps did (FinOps muscle built after BigQuery queries costing 30k). Levers: inference chips (Google TPUs), open-source models catching up (~9-month gap China/Anthropic), distilled models running on a laptop (Gemma 4, MoE 26B / 4B active, « designed in Paris »). A frontier model isn't needed for the group's ~300 digital products.
10. Model variation = a real production problem. when a model changes, retesting is required (different results); providers requantize / downgrade / route to lower capacity. They stopped upgrading once it was « good enough » (marketplace sorting: 15M products → 7M published, via OpenAI). The goal is « to be fast, but on one's own terms ».
11. Production awareness. Google has it while startups (OpenAI, Anthropic) don't yet (e.g. a bug where Codex/Claude limits were burned through in 1 hour instead of a full day, reset at 5pm).
12. His biggest concern: lock-in of the company's intelligence (the company's configuration baked into an agentic harness / « adeo.md »; « redoing the ERP of AI »). Countermeasures: standard Kubernetes (no proprietary Cloud Functions), regularly testing APIs with other hyperscalers, attention to memory management (cold/hot, context engineering, where it's stored, who owns it).
13. Missing open-source building block. agent orchestration (registry, lifecycle, permissions — « which agent is allowed to do what » —, sharing skills company-wide). n8n / MCP only cover a small part of it.
14. Company memory. (« company brain »): « when it's not logical, it's historical » — capturing the context behind decisions. E.g. Leroy Merlin doesn't sell mattresses because the company is run on revenue per square meter (a bed takes up too much space for what it brings in); this should go into the company brain so a future product manager doesn't re-pitch the idea.
15. RAG / access control. hard to know what the person asking the question is entitled to (multiple authorization levels) — an unresolved problem for enterprise chatbots.
16. AI stack. Vertex AI (GCP), model selection based on need. Always keep an eye on « the exit door » (sovereignty).
17. Advice to another CDO. transformation is bespoke (culture, available capital); identify the real competitive differentiator in the customer experience. If the board can't even articulate it → basic IT / operational maintenance. If it can articulate it → go all in, build something complete end-to-end, set the precedent. Understand the technology mainly « so as not to get fleeced by vendors and hype ».
18. Decathlon end-to-end RFID. a success because they manufacture a large share of their products themselves (chip embedded at the start of the industrial chain) all the way through to downstream security. Simple metric: time saved across the whole chain (staff + customer) = instant ROI. Adeo only manufactures half → getting all suppliers' industrial tooling to change is very complicated.