Agentic AI implementations break where several processes are made “smart” at once. In distribution, the first process needs three traits: its data already lives in invoices and campaigns, the decision repeats, and being wrong has a visible cost. Promotion planning usually has all three; “brand strategy” usually does not.
Score the process with a question
If you cannot phrase the question the way a sales manager would ask it in the morning (“which campaign had no effect?”), the process is not ready for an agent. Start the autonomy ladder on the lowest rung: the agent proposes, a human executes. Jumping to automatic execution before claim and evidence are stable skips governance.
A ninety-day bar
- One question family (promotion, for example) with a repeatable answer, not ten new dashboards.
- Every answer must carry claim, evidence, and risk; otherwise it is still a report.
- One real decision in the ninety days must be taken from the answer, or it was not a test.