AI in operations and supply chain, the cases that already work
Demand forecasting, inventory, purchasing and planning: where artificial intelligence improves operational decisions, and the conditions required for it to work.
Operations is where artificial intelligence has the longest track record and the most measurable results. This isn't novelty: models have been used for forecasting and optimization for decades. What changed is that access to that capability got cheap, and that unstructured information can now be used too.
The cases with proven results
Demand forecasting. The classic case, and still the highest impact. A forecast that incorporates seasonality, trends, promotions, holidays and external variables like weather typically beats manual assembly, especially across many products and locations.
Inventory optimization. From the forecast, setting reorder points and safety stock by product and location based on real variability and target service levels. This is where the most visible saving appears: less capital tied up at the same service level.
Purchasing planning. Suggesting what to buy, how much and when, using each supplier's actual lead times rather than the promised ones.
Predictive maintenance. Where sensors exist, anticipating equipment failures from operating patterns. The benefit isn't only repair cost — it's avoiding unplanned downtime.
Routing and logistics. Optimizing routes around time windows, capacity and constraints. Fuel and hours savings usually appear immediately.
Reading trade documents. Invoices, packing lists, bills of lading and declarations. High volume, variable formats, and manual work that consumes specialist time.
The conditions you have to meet
Before evaluating tools, check three things:
History and granularity. Forecasting needs enough history, ideally two or three years, at the level where decisions get made: product and location, not just family totals.
Clean master data. Consistent product codes, correct units of measure, clear equivalences. This is where most projects stall: the same product carrying three different codes ruins any model.
Real lead times. Not contractual ones — actual ones. If the system plans on 15 days and the supplier delivers in 28, safety stock will be wrong no matter how good the model is.
The most common mistake: optimizing one link
Improving the forecast without adjusting replenishment policies changes nothing. Optimizing inventory without factoring real supplier lead times creates stockouts. Improving routing without coordinating with order picking just moves the bottleneck.
That's why scope should be defined by looking at the whole flow, even when implementation happens in stages.
What to measure
The indicators aren't new; what changes is how often you can look at them:
- Forecast error by category, compared against the previous method.
- Service level: share of orders complete and on time.
- Days of inventory and capital tied up.
- Stockouts and estimated lost sales.
- Logistics cost per unit delivered.
The proof that a project works is two of these improving without the others getting worse. Cutting inventory at the expense of service isn't an improvement — it's moving the problem.
How to start
One practice that works well: run the new forecast in parallel with the current method for two or three complete cycles, without changing a single decision. Compare the error of both over the same periods.
If the new method wins consistently, adopt it for a subset of categories and expand from there. If it doesn't, the project cost three months of parallel calculation instead of a full transformation that was never going to work.
Frequently asked questions
How much historical data is needed to forecast demand with AI?
As a reference, two to three years of history at weekly granularity by product and location. With less history, traditional statistical methods are more stable and usually preferable.
Does AI replace the demand planner?
No. It produces the base forecast and removes manual assembly work, but the planner contributes context that isn't in the data: promotions, portfolio changes, market conditions and commercial decisions.
What concrete benefit can be expected in inventory?
In well-executed implementations, 10% to 25% reductions in tied-up stock while holding or improving service levels. The range depends on the starting point and data quality.