AI in inventory management for wholesale and food: a vendor-neutral step-by-step plan

What does AI automate in inventory management?

AI in inventory management is not a standalone system. It is three steps that follow on from one another. First, a demand forecast for each item, preferably with a range rather than a single number. Then the translation of that forecast into safety stock and ordering parameters. Finally, an order recommendation that the planner assesses. What you automate is the arithmetic: running every item through the same calculation each week. You do not automate the decision, or the knowledge that is not in the data.

Microsoft describes what that looks like in a mainstream ERP in its public documentation on the Sales and Inventory Forecast extension for Business Central. The extension forecasts sales based on sales history, shows when an item is likely to sell out and helps to create replenishment orders for suppliers. The input consists of the posting date, the item number and the quantity from item ledger entries of the type Sale. The documentation also states the limitations. The forecast is aggregated across all locations and variants. And if there is too little data or the data varies widely, the service does not produce a forecast.

We are not citing this example as a recommendation. It does, however, show two things that apply to any tool. A forecast that runs on sales entries sees only what you sold, not what customers asked for. And every tool has a point at which the data is too thin or too erratic. Ask every vendor for the same openness.

Why inventory management is a good first AI process

Inventory comes with plenty of historical data and an outcome you can measure: lost sales, inventory value and, in food, waste. That is why it is one of the processes that are the first to lend themselves to AI in wholesale and logistics. In research by evofenedex (the Dutch association for trade and logistics businesses) into AI in trade and logistics, conducted among almost three hundred logistics professionals, digital assistants, inventory management and demand forecasting are cited as the most promising applications. The same research shows where the bottleneck lies. Only 42 percent of the companies have their internal data largely in order, and at a third of them it is in order only to a limited extent.

Moreover, AI is not yet routine in logistics. According to provisional figures from CBS (Statistics Netherlands) for 2025, 45 percent of companies with 50 to 250 persons employed use AI. Across all company sizes, 12 percent of companies in the trade sector use AI, and 7 percent in transport and storage. Of the companies that use AI, 4 percent use it for logistics purposes, the lowest share of all the purposes CBS lists.

Step 1: make the data fit for use, starting with sales and demand

A forecasting model is only as good as the history it learns from. Before a model comes into play, check these points:

• Sales are not the same as demand. If an item was sold out, your history shows the stock limit and not what customers asked for. Record and flag lost sales, or a model will learn last year's shortages as normal demand.
• Promotions, special offers and large one-off orders are labelled with date and type, so that an outlier is recognisable as an exception.
• The item master is correct: no duplicate numbers, up-to-date units and complete supplier data. Changed item numbers are linked to their predecessor; otherwise a replacement item starts with no history.
• Lead times per supplier are recorded as they really are, not as they were agreed. A forecast only becomes an order recommendation once you know how far ahead you need to look.
• Ordering terms are recorded: minimum order quantity, order multiple and fixed ordering days for each supplier.
• In food: batch, expiry date and waste by reason. Without a reason, you cannot see whether waste comes from over-buying or from a quality problem.
• Every data flow has a named owner. Without an owner, quality declines unnoticed after the pilot.

Step 2: classify the product range by value and predictability

Not every item warrants the same approach. A useful classification is ABC-XYZ analysis. ABC tells you how much an item contributes to turnover or inventory value. XYZ tells you how regular demand is. An A item with regular demand is where a forecasting model shows most quickly what it is worth: there is plenty of history, and an error there costs money.

Slow movers with many weeks without sales are a different matter. The open-access textbook Forecasting: Principles and Practice by Rob Hyndman and George Athanasopoulos explains that the methods in the book assume continuous values. For series with small counts, such as 0, 1 or 2 per period, other methods are needed. Croston's method is widely used for this, but according to the authors it is known to be biased, and there is no formula for calculating prediction intervals for it. For such items, start with a simple fixed ordering rule and let a model prove first that it is better.

New items have no history. For these, the planner designates a comparable item as a reference and records that judgement. That remains the way of working, even with AI.

Step 3: run the forecast alongside your existing ERP

AI inventory management does not require a new ERP. The sensible order is the other way round. Choose one clearly defined product group from the A category, have a model produce an order recommendation every week and set it beside the planner's own. Take a period with ordinary weeks and at least one busier spell. Only then do you decide on software, with evidence from your own operation rather than a demo.

Also test the model retrospectively against your own history. Hyndman and Athanasopoulos stress that how well a model fits its own training data is not a reliable measure of the actual forecast error. So hold back a period the model has not seen, and compare on that period only. In their textbook, that is typically about 20 percent of the available data.

Choose the time unit that fits your ordering rhythm. For its own extension, Microsoft advises forecasting by month, with a horizon of twelve months. If you order from a supplier weekly, you also need a weekly forecast, covering at least the lead time.

The starting point in the pilot: AI advises, the planner decides and signs. That makes the planner a co-owner of the system rather than its opponent.

Step 4: translate the forecast into safety stock and ordering parameters

A forecast is not yet an order. Between the two sit the safety stock, the reorder point and the terms agreed with the supplier. The safety stock absorbs the uncertainty in demand and in lead time. That is why a range around the forecast is more useful than a single number. According to Hyndman and Athanasopoulos, a point forecast without an accompanying interval can be of almost no value, because you cannot see how accurate it is.

The margin grows quickly as you ask for more certainty. In their standard table, for normally distributed errors, the limit of an 80 percent interval lies at 1.28 standard deviations on either side of the forecast, that of a 95 percent interval at 1.96 and that of a 99 percent interval at 2.58. Each extra percentage point of certainty therefore costs more margin at the top of the scale than at the bottom. The same logic applies to safety stock. So choose the desired level for each item group, not for the whole product range at once.

In its best practices for Business Central, Microsoft gives practical pointers that are also useful outside that package. Use safety stock for items with large fluctuations in demand and for items that are subject to service agreements. Base the reorder point on the demand profile: if average demand is 100 units during a lead time of seven days, the reorder point is at least 100. Allow a safety lead time of one to six days, and a longer one for as long as a new supplier's delivery performance is unknown.

Next come the terms agreed with the supplier. If the minimum order quantity is higher than what you sell within one lead time, that term determines your inventory more than the forecast does. A well-designed recommendation makes this visible, so that the buyer can negotiate on it.

In food, shelf life limits the buffer. Safety stock that sits longer than its remaining shelf life becomes waste. For each product group, record what an error really costs on the high side and on the low side, and base the recommendation on that rather than on the middle scenario. You can read more about waste and quality in our article on AI in food manufacturing.

Step 5: work by exception rather than item by item

The biggest change for the planner lies not in the forecast but in where the attention goes. Instead of checking every order line, the planner assesses the items that the system flags as exceptions. Record every deviation from the recommendation, with the reason. This is not to check up on planners, but because those adjustments show where the model is consistently missing something, and where a planner corrected on gut feeling when the model was right.

Agree in advance what counts as an exception, for example:

• The recommendation differs sharply from the previous order or from the item's own history.
• The expected inventory drops below the safety stock within the lead time.
• A supplier delivers later than the recorded lead time.
• A new item, a discontinued item or an item with too little history.
• In food: stock approaching the end of its shelf life, or a batch that has been blocked.
• A customer, promotion or closure day that the planner knows about and the model does not.

Step 6: evaluate after the peak

The busiest weeks in particular yield a lot of instructive data. So schedule a set evaluation after every peak; after the December peak, that means January. How to prepare for that peak itself is covered in our article on demand forecasting before the December peak. Feed the outcome back into the data: flag the lost sales and promotions from the peak, so that next year the model learns from actual demand. The evaluation itself centres on four questions:

• Where was the forecast furthest off, by week and by item?
• Could that deviation have been predicted at the time? Only that category is a point for improvement.
• Which of the planner's adjustments turned out well, and which did not?
• Do the safety stock, reorder point or exception rules need adjusting, item group by item group?

What should you measure in AI inventory management?

Record a baseline measurement before the pilot begins. Without a starting point, the evaluation becomes a matter of gut feeling. After that, always measure in the same way:

• The forecast error for each item group, measured on weeks the model has not seen. An absolute measure, such as the mean absolute error in units, is easy to read for an individual item.
• Be careful with percentage errors. According to Hyndman and Athanasopoulos, they become infinite or undefined when demand is zero and extreme when demand is close to zero, and they penalise negative errors more heavily than positive ones. For slow movers, that is a real problem. To compare items with one another, they put forward the scaled error (MASE) as an alternative.
• Bias: is the forecast systematically too high or too low? A small average error can hide a persistent overestimate in one group.
• Lost sales and service level for the chosen product group.
• The average inventory value and the inventory turnover, at the same service level.
• In food: waste by reason, in kilograms or euros.
• How often the planner deviates from the recommendation, and who turned out to be right.
• The time planners spend on routine checking, compared with the time spent assessing exceptions.

What remains human work?

The planner decides and signs; AI advises. That is not a formality. A model knows only what is in the history: not that a new customer has been won, that the buyer agreed a promotion last week or that a supplier is closed in the last week of the year. That knowledge sits with people. It belongs in the system as an input, not as a correction after the fact.

In food, there is the added responsibility for food safety. Article 18 of the General Food Law (Regulation (EC) No 178/2002) requires you to be able to identify who has supplied you with food and which businesses you have supplied. Which batch goes out first, when a batch is blocked and what happens in the event of a recall remain decisions made by people within your quality system. An order recommendation must not work around that.

Waste is also on the European agenda. Directive (EU) 2025/1892 obliges the Member States to take measures to reduce food waste in processing and manufacturing by 10 percent by 31 December 2030, compared with the annual average for 2021 to 2023 inclusive. That target applies to the Member States, not directly to your company.

The people who work with the model must also understand it. Article 4 of the EU AI Act (Regulation (EU) 2024/1689) has applied since 2 February 2025. Since the amendment by Regulation (EU) 2026/1744, in force from 27 July 2026, it has asked providers and deployers of AI systems to take measures that support the development of AI literacy among their staff and others who work with those systems on their behalf. They do not have to guarantee a specific level. For planners, that comes down to knowing what a forecast is based on, recognising when a model is working outside its range and being able to explain why they followed or adjusted a recommendation. You can read more about this in our article on AI literacy.

How do you test a vendor or tool?

A savings percentage quoted upfront tells you little, because the result depends on your product range, your demand pattern and your data. Even a good trial is no guarantee: our article on AI projects that stall after the pilot explains why AI projects can stall after the pilot and how to provide for that in advance.

So ask a vendor these questions instead:

• Will you test the model on a period from our own history that it has not seen, and set the outcome against actual sales?
• How does the model deal with lost sales and sold-out periods?
• What happens with items whose history is too short or too erratic?
• Does it forecast by location and by variant, or in aggregate?
• Does it give a range, and can we set the desired level of certainty for each item group?
• Does the system record who adjusted a recommendation, and why?
• Can we take our data and settings with us if we switch?

What you can realistically expect

What is achievable depends on the breadth of your product range, how erratic demand is and how good a state your data is in. A reasonable goal after a good pilot: fewer lost sales in the chosen product group, a lower average inventory at the same service level, less waste in food, and planners who shift their time from checking to assessing. How much that amounts to is shown by your own baseline measurement, not by a percentage from a brochure.

On a Practice Day, we work with the people who run the process every day on one work process from your own operation, for example drawing up and assessing the order recommendation for one product group. The team learns it while it is being built. The weeks in which the recommendation then runs alongside the planner fall outside that one day. If the process touches a standard such as HACCP, BRC or IFS, we record the working agreements in the language of that standard. A Practice Day costs from €2,500 per group, excluding VAT. We do not promise a saving upfront.