AI in Operations and Supply Chain

When executives at Procter & Gamble looked closely at the order patterns for Pampers, one of their best-selling products, they found something odd. Babies use diapers at a fairly steady rate, and sales in retail stores did fluctuate, but not by much. The orders that distributors sent to P&G were another matter: the executives were surprised by how much they swung. And when they looked at P&G’s own orders for materials from suppliers such as 3M, the swings were greater still. Hewlett-Packard found the same shape in one of its printers: modest swings in a reseller’s sales, bigger swings in the reseller’s orders, and the biggest in the orders from the printer division to the company’s own chip division¹.

Nobody in that chain lacked a forecast. Each tier forecast diligently, from the orders that arrived at its door. Hau Lee, V. Padmanabhan and Seungjin Whang named the result the bullwhip effect, and their diagnosis still reads as a warning for anyone buying AI for operations: the problem was not the quality of any one prediction. It was how decisions were connected, how often they were made, and what information each decision maker was allowed to see.

The value is in the physical move

Most of the AI this course has discussed so far ends in words: a draft, an answer, a summary. Operational AI ends in the physical world. A truck leaves or waits. A part is ordered or not. A machine is stopped tonight or runs until it breaks. Stock sits in one warehouse while the shortage is in another. That changes how value is created, and it changes what can go wrong.

The general distinction between predicting, forecasting and optimizing has its own chapter, AI in Prediction, Forecasting and Optimization, and AI in Finance already made the point that a forecast is judged by the decision it changes. Operations adds three hard conditions. The signal has to reach someone who is allowed to act. It has to arrive in time for the action to matter. And the action has to happen inside a chain of other decisions that can either absorb it or amplify it.

Organizations tend to spend on the first two steps of this loop, because that is where the technology is. The loop often breaks at the third and fourth, because that is where the organization is. The executive test for any operational AI proposal is therefore blunt: which physical action will happen differently, who will take it, and how fast?

Operations is a set of coupled decisions

It helps to see operations not as a list of AI tools but as a family of decisions. Planning decides what demand to expect and how supply should follow. Purchasing decides what to buy, from whom and when. Production decides what to make and in what sequence. Maintenance decides when to intervene on an asset. Quality decides what passes inspection, and camera-based inspection can in principle check every unit on a line rather than a sample, a capability whose payoff depends on what the line does with each rejection. Logistics decides routes, loads and delivery slots.

The word that matters is coupled. Each of these decisions feeds the next, so a local improvement can make the whole worse. Lee and his colleagues traced the bullwhip to four causes, and every one of them is a decision rule, not a data problem¹. Two of them should worry any buyer of AI forecasting. If each tier feeds its model the orders it receives rather than what end customers actually consume, a faster, more responsive model will pass on the swings faster. And if orders are still batched once a month, a daily forecast changes nothing until the batch is placed.

The remedies Lee and his colleagues proposed were organizational: share sell-through data with suppliers, order in smaller and more frequent batches, stabilize prices, and allocate scarce supply by past sales rather than by what customers claim to need. AI makes several of those remedies cheaper to run. It does not make them happen.

Every decision has a clock

Operational decisions run on very different clocks. A vibration reading on a pump may need an answer in minutes. A delivery route is re-sequenced within the day. Replenishment runs daily or weekly. The production plan is often fixed a week or a month ahead. Supplier contracts and capacity move over quarters. An AI system is worth what the fastest decision it can actually change is worth.

Suppose, as an illustration, that a logistics team receives an early and correct AI warning that a storm will close a key port on Friday. The alert lands in a shared inbox on Tuesday. On Wednesday everyone has seen it and no one owns it. On Thursday someone asks who may reroute the containers, and the answer involves three approvals and a carrier contract nobody has read. On Friday the port closes and the containers wait. The prediction worked. The decision clock did not.

Two practical consequences follow. First, real-time systems are worth their cost only where the decision is real-time; a weekly replenishment decision does not need streaming data. Second, the cheapest improvement is often to shorten the decision clock itself: move from monthly to weekly ordering, give a planner authority to transfer stock between warehouses, pre-approve the alternative carrier. Those are management decisions, and they are often worth more than the next model upgrade. Route optimization shows the scale of small daily moves: UPS expected its ORION planner, which chooses the order of a driver’s stops and lets drivers depart from the planned sequence where they know better, to cut about 100 million miles a year at full deployment, an expectation rather than an audited result².

Maintenance: score the downtime, not the model

Predictive maintenance replaces two old habits: servicing on a fixed calendar, and running equipment until it fails followed by an emergency repair. Condition-based maintenance uses sensor data, such as vibration, temperature and oil analysis, to estimate when an asset is likely to fail and to choose when to intervene³. The promise is real. A US Department of Energy guide estimated that a working predictive program saves 8 to 12 percent over preventive maintenance alone, and cited industry surveys reporting 35 to 45 percent less downtime and 70 to 75 percent fewer breakdowns⁴. Those are survey averages from 2010, before modern machine learning; treat them as the size of the prize, not as a forecast for your plant.

The trap is to measure the model instead of the maintenance. Every alert sends a technician, and many stop a machine. A model that catches more failures by raising more alarms can cost more than it saves.

The figure is an illustration, not plant data, and so is its arithmetic. Suppose each failure caught early avoids 40,000 of unplanned downtime, and each false alarm costs 7,500 in technician time and a planned stop. A sensitive setting flags 100 machines in a quarter and catches 20 real failures: 800,000 saved, minus 80 false alarms costing 600,000, leaves 200,000. A stricter setting flags 40 machines and catches 16: 640,000 saved, minus 24 false alarms costing 180,000, leaves 460,000. The second model catches fewer failures and creates more than twice the value. The general logic of unequal error costs belongs to AI in Prediction, Forecasting and Optimization; the operational lesson is to report downtime, maintenance cost, technician hours and lost production, and to count the work every false alarm creates. Teams that are sent to too many healthy machines also stop trusting the alerts, which is a cost no dashboard shows.

Authority grows with physical consequence

Because operational AI moves physical things, who may act matters more here than in most functions. The useful test is reversibility: can the action be undone before it costs anything? A purchase order can usually be cancelled the same day. A container that has sailed, a batch that has been released or a line that has been stopped cannot.

At the bottom, AI acts within limits on routine, reversible actions, and every action is logged. In the middle, it recommends and a named person approves. At the top, people decide, protected by deterministic safeguards such as interlocks and certified control systems that do not depend on a model being right. The tiers are not fixed: an action can move down a tier once its record is good, and should move up after an incident. The general design of human oversight is covered in Module 06; here the point is that in operations the tier is set by physics, not by enthusiasm.

Story: the stores that ignored the buyers

An automobile spare-parts retailer had given its merchants, the buyers who decide what each store stocks, a data-driven tool. It drew on local car registrations, weather and each product’s sales history to recommend stocking decisions. Like many firms, the retailer let the merchants override the tool, on the reasonable grounds that they might know things the tool did not. Saravanan Kesavan and Tarun Kushwaha ran a field experiment to find out what that discretion was worth. For twelve months, across more than 30,000 products, the merchants’ overrides were carried out in some stores and ignored in a random selection of others⁸.

Overall, the stores that honored the overrides were 5.77 percent less profitable than the stores that followed the tool to the letter⁸. That sounds like a verdict against human judgment, and it is not. The effect reversed with the product’s life cycle. For growth-stage products, the overrides drove more than 23 percent greater profitability; for mature and declining products, the merchants did worse than the tool⁹⁸. The explanation is intuitive once stated: for a new part, the tool has little history to learn from, and an experienced buyer who knows which car models are arriving in the area does have an edge. For a part that has sold steadily for years, the tool has seen everything the buyer has, and the buyer’s adjustments mostly add noise. The two figures measure different things: the 5.77 percent compares all products in the two groups of stores, while the 23 percent covers only the growth-stage subset.

A later experiment at a spare-parts retailer, by the same two researchers and Dayton Steele, tested a different design. Instead of overriding the tool’s output, staff could adjust the inputs to the forecast that fed an inventory algorithm. Those adjustments raised profitability by 4.92 percent on average compared with automation alone, and the authors note that forecast performance need not translate into profit performance¹⁰.

Read the two results together and the lesson for an operations leader is not “trust the algorithm” or “trust your people”. It is that the override is a design decision. Where may people intervene: on the inputs, where they add what the model cannot see, or on the final answer, where they tend to second-guess what it already knows? For which products, suppliers or situations? And is each override recorded and scored in profit, so the organization learns whose judgment adds value and where? An override button is not an override policy.

What this means for leaders

Operations is where AI’s predictions turn into trucks, parts and machine hours, and where a good signal is most easily wasted. The leadership work sits in four decisions. Map the decisions that the AI will feed, and check that improving one will not amplify swings in the next. Set the decision clock deliberately, and shorten it where the money is. Design the override: where people may intervene, for which cases, and how each intervention is scored. And set authority by physical consequence, writing down what the system may do alone, what it may only recommend, and what stays with people and hard safeguards. Through all of it, insist on an operational scoreboard: service level, stockouts, expired or obsolete stock, downtime, maintenance cost, miles, fuel and working capital.

Check yourself

1. The bullwhip effect is mainly caused by poor forecasting models.MythFact
2. A maintenance model that catches more failures always creates more value.MythFact
3. Real-time AI is worth its cost only where the decision itself is made in real time.MythFact
4. In a randomized field experiment, buyers’ overrides of a stocking tool lowered profit overall but raised it for new products.MythFact
5. The EU Machinery Regulation, applying from January 2027, covers machinery safety components that use AI.MythFact

Reflection: follow one signal

What comes next

Operations shows AI at the physical core of the enterprise, where a prediction becomes a truck, a part or a stopped machine. The next chapter turns to a function where AI helps build the systems that run the business itself, and where the same question follows us: what actually changes when the AI gets better? That chapter is AI in Software Engineering.