Short answer: inventory suits AI because every useful question here requires combining four or five sources — stock, open purchase orders, open sales orders, lead times, consumption history. That join takes a person twenty minutes per item, and gets done only for the ten items somebody happened to think about.
Coverage, not stock on hand
"How much do we have" has a simple answer and almost no value. The question that matters is whether what you have, plus what is arriving, covers what is already promised and what you expect to sell before the next receipt.
An assistant splits shortage into three different situations, and that split is what makes the list actionable:
- No purchase order exists at all — needs placing, and the sooner the better.
- An order exists but arrives later than needed — needs expediting or an alternative.
- The item is already promised on open sales orders — the expensive category, because this is a missed commitment rather than an empty shelf.
In a standard report all three look identical: an item below its reorder point. The required actions are different.
The supplier who is always late
Planning runs on confirmed dates. If a supplier consistently arrives eleven days late, the plan fails quietly, every time, and without a visible cause.
The history of actual receipts against confirmed dates sits in Business Central in full. Computing the average deviation per supplier is minutes of work for a system and days for a person, which is why nobody does it. An assistant can substitute real lead times for nominal ones and show immediately which items move into the risk zone as a result.
The underlying calculation: reorder point and safety stock.
Capital that is not moving
The mirror problem, and usually the larger number. Stock that does not move is invisible on a P&L and painful on a balance sheet. Business Central holds everything needed to quantify it: last movement date, quantity, value and turnover per item.
What makes the output usable is ranking by money rather than by count. Four hundred slow-moving SKUs is a report nobody acts on. "€180,000 has not moved in a year, and 60% of it sits in eleven items" is a decision you can make today. The follow-up questions — is it obsolete, seasonal, or tied to a contract — are exactly what an assistant can answer in the same conversation.
AI and the planning worksheet
Business Central has planning and requisition worksheets, and they work. Worth stating plainly, because this is usually sold the other way round: AI does not replace the planning engine.
What AI adds is the reasoning around it: explaining why a suggestion appeared, flagging when it rests on unusual demand history such as a one-off large order, and combining the recommendation with actual supplier lead times rather than the configured value.
The flag for one-off demand deserves emphasis. The planning engine does not distinguish regular demand from a single spike: if a customer took a year's volume in one delivery, next month the system will propose buying the same again. That is one of the most common causes of excess stock, and it is visible in the demand structure.
Conversely, if item cards are not maintained, AI inherits the drift. Neither situation improves by treating the model as a substitute for maintained master data.
Where inventory analysis gets subtle
Two things that break the numbers quietly.
Costing method. FIFO, average and standard behave differently, and value pulled before cost adjustment has run is provisional. The assistant should report that caveat rather than present a provisional figure as final. Mechanics: inventory costing methods.
Locations. Stock that exists in the company but sits in the wrong warehouse does not exist for a specific order. Analysis at company level where the business runs at location level is technically correct and operationally useless.
A third, more fundamental one: if book quantities disagree with physical stock, analysis on top of them is planning imaginary goods. Causes: why stock figures do not reconcile.
Where to start
The cheapest first step is not a project but a check. Take twenty items that cause the most trouble and work out by hand how many are currently at risk using actual rather than confirmed lead times.
If the result surprises you, you have a reason to compute it regularly. If it does not, your setup is in good shape and AI will add little here — which is a useful and inexpensive thing to learn.
What this looks like as a service: AI for inventory in Business Central.
From analysis to action
Everything above is read-only. An AI agent extends it: creating purchase orders from approved suggestions, updating reorder points based on observed demand, or opening a supplier query when receipts start slipping.
Because an unposted purchase order is recoverable, this is one of the areas where unattended creation with review before release is a reasonable starting position — more so than in finance, where mistakes are costlier to unwind. Where exactly the line falls is decided per process, but the asymmetry between the two areas is real and worth using.
Frequently asked questions
How is this different from the planning worksheet?
The worksheet computes suggestions from your item setup and does it well. AI explains why a suggestion appeared, flags when it rests on a one-off demand spike, and substitutes actual supplier lead times for configured ones. It is an addition, not a replacement.
What if our item setup is out of date?
Fix it first. Reorder points, lead times and order modifiers have to reflect reality, otherwise both the planning engine and AI calculate from wrong inputs — and AI will do it faster and at greater scale.
Does it account for seasonality?
It can identify seasonal patterns where enough history exists — typically two to three years. With less, it will describe a coincidence as a trend. The assistant should state how many periods the conclusion rests on.
Does it handle multiple locations?
Yes, and it should. Company-level analysis hides the case where stock exists but sits in the wrong warehouse for a given order. Location-level analysis needs your location setup to match how you actually operate.
Will inventory values match our balance sheet?
Only if the analysis respects your costing method and cost adjustment timing. Figures pulled before cost adjustment has run are provisional, and an assistant should flag that rather than present them as final.
Can an agent create purchase orders automatically?
Within its permission set, yes. Because an unposted purchase order is recoverable, creating them unattended with review before release is a reasonable starting position — more so than in finance.