An AI replenishment agent runs your inventory for 26 weeks. You don't place a single order - you set its policy: how autonomous it is, how cautious, how reactive. It's the same governance question as the lab's live AI Assistant: how much do you let the machine decide alone?
Agentic AI is not just a model that predicts - it is a policy wrapper that decides WHEN the machine may act alone and when a human must sign off. You never place an order here; you govern the thing that does. The sliders are governance, not forecasting.
A base-stock agent forecasts demand (weight α), adds a safety buffer (z·σ) and tops inventory up to S every week. The approval threshold is the autonomy dial: any order above it queues for human sign-off and lands a week late - safe-feeling, but slow.
Approval latency during a promo costs more than the risk it prevents: a delayed big order becomes a $12/unit stockout. Govern by exception size, not by distrust - raise the threshold so routine big orders flow, and keep sign-off for genuine outliers.
