AI in supply chain planning is only as good as the plan underneath it

AI in supply chain planning means using models that learn from data, or search through huge numbers of options, to help with specific planning decisions such us how much demand to expect, how much stock to hold and where, and what to order from which supplier and when. I find it more useful to define it by the decisions it improves than by the technology, and it only pays off when those decisions already rest on data and processes that work.

In this piece I want to look at what AI really does in a planning process, how the pieces fit together, and what has to be in place before it shows up in the results.

What AI actually does in planning

Rather than treating AI as one thing, I think it’s more helpful to look at the decisions it supports. In most planning teams these fall into three groups.

• Demand. Estimating future demand by item, customer and location, spotting when a pattern has really changed rather than just wobbled, and telling a one-off order apart from a new trend.
• Inventory. Grouping items by how they behave, setting service levels and safety stock for each group, and deciding where stock should sit across the network.
• Supply. When to order and how much, how to react when a supplier’s lead times or reliability change, and whether to expedite, rebalance stock or live with a short-term service risk.

If someone pitches you an AI capability and can’t tie it to one of these decisions, and to the person who makes it, I’d be sceptical that it will change much.

How it works, from data to decision

Any useful application of AI in planning follows roughly the same chain. Breaking it down helps show where a project is likely to get stuck.

1. Data and signals. Sales and order history, stock positions, open orders, supplier confirmations, promotions and other known events. Clean data beats lots of data.
2. Prediction or optimisation. A model either estimates what’s likely to happen, or works out the best response within the business’s constraints. Often it does both, in that order.
3. Recommendation or alert. The result comes out as a suggested action, or as an exception someone needs to look at, with the reasoning shown.
4. Planner review. Routine recommendations inside agreed tolerances are accepted, either automatically or in bulk. Exceptions go to a planner, who can accept, change or reject them and note why.
5. Execution and measurement. The decision is carried out and the outcome compared with what was expected. That feeds back into the data, and into the rules about what can be automated.

When AI projects disappoint, step two is rarely the problem. More often it’s step one, because the data can’t be trusted, or step four, because planners can’t see why the system recommended something, so they work around it in a spreadsheet.

Start with the decision

Donato Masi and colleagues have looked at AI-driven dynamic pricing for high-value products that have to be sold within a fixed window (Chen, Hu, Qi and Masi, 2026). I’m less interested in the pricing itself than in how the model is set up. It uses its view of where offered prices are heading to set the lowest price it should accept at each point. If prices look set to rise, it holds off. If they look set to fall, it adjusts its threshold as the deadline gets closer. The forecast is only useful because it’s tied to a decision with a deadline.

Planning teams often do it the other way round. They compare forecast accuracy across models, pick the one with the lowest error and assume the business will benefit. But a slightly more accurate forecast on items nobody acts on is worth less than a rougher one that tells you three weeks earlier to pull forward a purchase order or move stock between sites. What counts is what you’d do differently because of it.

What needs to be in place first

Visibility first

One of the most useful studies I’ve read on this doesn’t involve AI at all. Kamble, Clegg, Masi and Al-Bazi (2024) worked with a small, traditional manufacturer that kept its production records on paper. Instead of replacing an old hand press, they fitted it with a cheap proximity sensor and microcontroller to log every cycle and every stoppage, and fed the data into a dashboard. For the first time, the business could plan from what was actually happening on the shop floor rather than what people remembered.

The know-how was already there; what was missing was a clear view of what was going on. Much larger organisations have the same problem, just with more sophisticated paper: lead times nobody has checked since the supplier changed, stock records that slowly drift, and demand history distorted by stockouts, so that what looks like weak demand is really poor availability. Point an AI model at that and it will learn your mistakes and repeat them confidently. As Professor Masi puts it:

“AI doesn’t remove the need for good data. It raises the price of bad data.”

For some businesses, the best use of next year’s AI budget would be the groundwork underneath it: tidying master data, keeping lead times up to date and recording true demand, including sales lost to stockouts.

Not every item behaves the same way

Segmentation has long been at the heart of Slimstock’s approach to inventory planning, and Masi’s research arrives at the same place from a different direction. Working with a utility company, he and his co-authors showed that infrastructure projects should be prioritised by their demand profile rather than just their type, and that repeatable work gets more efficient when it’s treated as repeatable (Masi et al., 2018). The same applies to inventory. A fast-moving consumable, a seasonal line and a spare part that sells twice a year need more than different forecasts. They need different service targets, different review cycles and different levels of attention from planners.

It’s tempting to run one sophisticated model across the whole range and let it sort things out. But however clever the model, that’s still a one-size-fits-all policy. Where AI could really help is in reclassifying items every planning cycle, spotting when one moves into a different segment, and changing the policy as well as the number.

Process and ownership

Every recommendation needs an owner, a rule for when it can be accepted automatically, and a clear route for exceptions. Without those, AI just adds a second opinion to a process that already struggles to act on the first one.

Keeping people in the loop

The people side is usually what decides whether any of this sticks. In a related study of a small manufacturer’s digital transformation, again with Masi among the authors, the change was handled as a shift in people and processes as much as in systems (Panchal et al., 2024).

This matters most in planning. A planner with fifteen years’ experience won’t hand decisions over to a system they can’t question, and they shouldn’t. If people can’t see why the system recommends something, they won’t use it. They’ll override it in a spreadsheet, and the business ends up paying for two planning processes.

Where automation is appropriate

Routine replenishment of stable, high-volume items, with reliable data and recommendations inside agreed tolerances, is exactly the sort of work a system should take on. Automating it gives planners more time for the decisions where they make a difference.

Where expert judgement is needed

New products with no sales history, promotions or events the system hasn’t seen before, supplier problems, expensive or long-lead-time items, and cases where the signals conflict. These should go to a planner deliberately, as part of how the process is designed.

How planners challenge and override

• Show the drivers. Every recommendation should show what changed and why, whether that’s the demand assumption, the lead time or the constraint that’s holding things up.
• Ask for a reason. Overrides should be quick to make, with a short reason code so everyone can learn from them.
• Track the overrides. Check whether overrides make outcomes better or worse. Good ones should feed back into the model and the rules; poor ones show where training or trust needs work.
• Move the line over time. As evidence builds, more types of decision can move from review to automatic acceptance, and some may need to move back.

So the planner’s job changes, with less time spent calculating and more spent questioning. The work that’s left, like judging a promotion the system has never seen, working with a struggling supplier, or challenging an assumption in sales and operations planning (S&OP), is where good planners matter most.

AI-assisted and autonomous planning are different things

These two terms often get mixed up. In AI-assisted planning, the system spots problems, makes predictions, tests options and recommends what to do, while people set the policies, deal with exceptions and stay accountable for the results. In autonomous planning, the system carries out decisions from start to finish without anyone reviewing them.

Autonomy can work for narrow, well-understood decisions with reliable data, clear limits and good measurement, like the routine replenishment above. Applied to a whole planning process, it strips out the context and accountability good planning relies on. And no model will make up for poor data or a broken process.

Three questions to ask first

So next time AI in planning comes up, I’d suggest leaving the technology until later and starting here:

1. Which decision will this change, and who owns it?
2. Is the data underneath it a fact or a habit?
3. Will my planners understand why it recommends what it does, and can they challenge it?

If you can answer all three, AI can make a real difference to how you plan. If you can’t, start by fixing the data and the process. It’s less exciting than a new model, but it’s what makes the model worth having.