Asking an AI tool to suggest reorder quantities

Give an AI tool a small table and clear rules, and ask it to show its working. Then check one row yourself before anything gets ordered.

What to put in the request

A chat tool knows nothing about your shop. Everything it needs has to be in the message. Include:

• A table with one row per product: name, code, on-hand count, and sales per week for the last several weeks.
• The supplier lead time for each product, in days.
• How often you order. Weekly, fortnightly and monthly orders need different cover.
• Your rules: a minimum order size, a case size, a cap on what you'll spend.
• Anything odd in the history, such as a closed week or a bulk sale to one customer.

Keep the table short at first. Five to ten products is enough to see whether the tool understands the job. The data-prep page lists what to tidy.

Ask for the working, not just the answer

A plain request gives a number and nothing else. Try this:

Work out average weekly sales for each product and show the sum. Suggest an order that covers the lead time plus a stated cushion. Round up to whole cases. List any product where you had to assume something.

The last sentence matters. It makes the tool tell you where it guessed. You can then fix the guess instead of discovering it later.

Check the numbers

A tool may state something false as though it were fact, which people call an AI hallucination.¹ In a table of figures, that could be a total that doesn't add up or a lead time you never gave. Check these things:

1. The inputs. Does each product's sales history match what you pasted? Copy errors happen.
2. The averages. Add up one product's weekly sales and divide by the number of weeks. Compare.
3. The cover. The usual rule multiplies average daily usage by the lead time in days.² Use the supplier's real wait from order to delivery.² Do the sum for one product and see whether the tool's figure covers it.
4. The cushion. Is it stated? Does it fit how much you can afford to tie up?
5. The rounding. Does the order match your case sizes and minimums?

The formula is set out in full on the reorder point page. If one row is wrong, assume others are and recheck the rest.

Test it against the past

You can test a tool on history you already know. Give it sales up to a chosen week and ask what it expects afterward. Then compare with what happened. Microsoft's forecast help suggests the same idea. Start the forecast before the last historical point, and you can compare the forecast to the actual data.³

Mind the layout of your data

If you use a spreadsheet forecast function instead of a chat tool, the layout rules are strict. Microsoft's FORECAST.ETS needs the timeline organized with a constant step between the different points.⁴ Weekly totals work. A list of random sale dates doesn't.

Keep the decision

Treat the suggestion as a draft order. You've counted the shelf, you know a supplier is changing prices, and you know the shop is closing early for a holiday. The tool doesn't. Approve the order yourself, and change any line that doesn't fit what you know.

Save the request you used and the answer you got, with the date. A month later you can see how close it was. The page on slow and fast sellers uses a similar approach for sorting products.