The autonomy ladder says which rung an AI agent may climb in FMCG distribution operations: propose, draft an action, execute behind an approval gate, execute inside a closed range. AI governance means those rungs are written in advance, not rewritten overnight after one “good” answer. Didrah stays on the propose rung. Jumping to automatic execution on stock, price, or dropping a customer skips human control. A supply chain is not a place where one correct answer licenses moving money.
What the rungs look like in distribution practice
Propose: a claim and a recommended action, a human executes. Draft: the order text or beat change is ready, nothing is sent. Approval gate: execution only after a named role signs. Closed range: for example a low-stock reminder, not a credit-limit change. Human-in-the-loop is the lower rungs. Human-on-the-loop is a rung that executes and a human can still cut. Human-out-of-the-loop has no place in distribution operations where money or stock moves.
Which operating decisions the agent must not own
Cutting a wholesaler, changing a contract price, zeroing a brand in one warehouse, and publishing a figure a policy check rejected. Those need a kill switch: if the agent approaches them, the answer should stop. A guardrail is this list, not a moral slogan. If the claim is right and the action is dangerous, the action still does not run. A correct claim is not a licence to execute.
When a rung may move up one step
When a golden set of questions is stable, traceability from claim to invoice is complete, and several cycles of human decisions used that same answer shape. Even then, raise the rung for one question family, not for “all of AI.” Starting with one process is this point. If the agent eval on that family weakens, drop the rung. The ladder is not one-way.
What to check on every distribution-operations agent answer
- Which rung the recommended action sits on; if it is unnamed, treat it as propose.
- Whether the action approaches stock, price, or dropping a customer; if yes, it stops.
- The audit log should say which policy allowed a figure to be published.
- Do not edit a rejected answer until it is “presentable.” Rejected means rejected.
Where AI governance shows up in weekly distribution operations
Start every answer with one rung line: it is a proposal, or it is close to execution. If it is close to execution and it touches stock or price, stop there. That habit beats a long policy document, because it repeats every week. The AI agent should write the rung in the structured output, not in a footnote. If a month passes and no answer was rejected, either the questions are too safe or the guardrail is off. Both are dangerous in distribution operations.