Before a manufacturer buys a new AI planning tool, it should record for one quarter what its planners do with the exceptions its current planning system flags, then start AI on the kinds of exception that cost the most.
Exceptions are the alerts a planning system raises when the plan will not hold, whether it is an advanced planning system, or APS, or the MRP run inside the ERP: a purchase order due after the material is needed, a line loaded past its capacity, stock below its safety level. BCG's report on supply chain planning, published in February 2026, finds that "APS platforms are widely embedded in most large organizations, yet process redesign and operating model changes frequently lag behind system deployment." As a result, BCG writes, many companies with modern tools use too little of what those tools can do and miss the benefits they expected. The same report finds that most value today comes from foundational applications, and it names exception management among them, next to forecasting, data interpretation and workflow automation.
Bain's 2026 CEO survey, published in July, asked 100 CEOs, only 21 of them at industrial companies, so its industrial figures rest on a small group: 90 percent of those 21 believe their AI programs are underdelivering, citing capability gaps, pilots that fail to scale and unproven returns. BCG's report adds that organizations that try to leapfrog planning maturity with AI alone tend to struggle, while those that layer AI deliberately onto stable planning foundations see more durable gains.
The record is kept where planners already work the exceptions, at their desks through the week, since a planning run flags far more of them than any meeting could discuss. Each week the list is ranked by the value of the customer orders each exception puts at risk. For the top 25, the planner notes whether they acted, what they did, such as expedite, reschedule or accept the delay, and who decided, and a random sample of 25 more from the rest of the list gets the same notes. The weekly planning meeting still reviews only the exceptions escalated to it. Once each noted exception's need date has passed, the planning manager checks it against what happened: whether the late order held up production, whether the overloaded line missed its dates, whether the shortage reached a customer.
For a maker of industrial pumps with plants in the US and Mexico, one line of the record might read: exception, castings from the Mexico plant due at the US machining line after the schedule needs them; action, expedited with premium freight across the border; decided by, the materials planner, with the plant manager approving the freight; result, a week later, the castings arrived and machining held its date.
After thirteen weeks of results, the record shows which kinds of exception the planners act on, which they leave, what the ones they left cost later in premium freight and overtime, and the value of the orders that shipped late because of them, while the random sample shows whether ranking by order value missed exceptions that turned out costly. The costliest kinds are the first places to put AI, working from the exception data the planning system already holds, whether the AI comes from that system's vendor or from an application built on its data: ranking each week's exceptions by what the record shows they cost when left alone, and drafting a response for a planner to approve. The same record gives any vendor of a new AI planning tool the list of decisions its tool would have to improve.
