Where does AI improve a supply chain first?
In the forecasts and the follow-through. Prediction is the classic win: demand, lead times, stockouts, supplier risk, all learnable from history your ERP already holds. The newer win is the follow-through, where an AI agent chases order confirmations, reconciles paperwork and drafts the replan when a shipment slips, work that is coordination rather than math. The forecasting improves decisions; the agentic workflow removes the clerical lag between them.
What data does supply chain AI need before it works?
Less than vendors imply and more than most ERPs cleanly hold. Forecasting needs order history with the dates things actually happened, not the ones backfilled at month end; supplier scoring needs delivery and quality records tied to the right supplier; an agent needs access to the inboxes and portals where the chasing happens. The honest first project is often data readiness work in disguise: if the ERP cannot say what happened last quarter, no model can say what happens next.
Which AI is best for a supply chain?
The question hides three different purchases. Forecasting and inventory intelligence usually ships inside the planning suite already in place. Coordination work arrives as an agent layer over email, ERP and supplier portals. Network and route optimization prices out at fleet scale. Judge each candidate against the one failure that costs the most money, with an ordinary AI vendor assessment rather than a feature checklist, because every vendor in this market demos well.
Where do robots and physical automation fit in?
At the points where the plan meets actual goods. Warehouse robots, automated picking and yard vehicles execute what the planning layer decides, and they belong to physical AI, where the economics reward volume and punish false starts. The production side of the same network has its own patterns under AI in manufacturing. For most mid-sized operators the software layer pays back years before the steel does, which is why the sequencing below starts with data rather than machines.
How should a mid-sized company start?
Not with a platform. Pick one measurable pain, stockouts on a product family, expediting hours, invoice mismatches, and fit the smallest tool that moves it, often plain workflow automation before anything learns. The supply chain rewards boring first steps, because every step feeds the next one data. The transport leg of the same story has its own economics and its own entry, AI in logistics; the sequencing discipline is the same AI adoption work as anywhere else in the company.