How to Use AI to Plan Stock for a Busy Week
Use AI to plan stock for a busy week by giving it your sales history, not your guesses. The exact numbers to supply and the prompt that works.
You can use AI to plan stock for a busy week, and it will do a decent job, on one condition: it must never be the thing estimating demand. Give it your numbers and it will do arithmetic you would otherwise do badly at eleven at night. Ask it how much you will sell and it will invent a plausible figure, because inventing plausible figures is what it does when it has nothing.
That distinction is the whole article. The rest is how to supply the numbers.
What to export before you start
Pull from your till, shop platform or spreadsheet. Five columns, nothing clever:
Item name
Units sold per day, for the last four to six weeks
Units sold during the equivalent busy period last year, if you have it
Current stock on hand
Supplier lead time in days, per supplier
If you have no sales history at all, stop. No tool fixes that, and a model asked to work without it will produce confident numbers with nothing behind them. Record a few weeks of sales first.
The prompt
Paste the data and be explicit about the boundary:
Here is my sales data for the last six weeks, plus the same week
last year, current stock and supplier lead times.
Next week is [event / holiday / local festival], which last year
increased overall takings by about [X]%.
Work out, per item:
1. Expected units for next week, based ONLY on the figures I gave you.
Show the calculation.
2. Units I need to order, given current stock.
3. The order deadline, given each supplier's lead time.
Rules:
- Do not estimate any number I have not supplied. If something is
missing, list it as a question instead of filling it in.
- Flag any item where last year's figure and the recent trend
disagree by more than 30%.
- Show the arithmetic so I can check it.The two rules at the bottom are the ones that matter. The first stops invention. The second catches the items where the data is genuinely ambiguous and a human needs to decide.
What good output looks like
Item | Recent daily avg | Last year's week | Expected next week | In stock | Order | Deadline |
|---|---|---|---|---|---|---|
Oat milk | 14 | 132 | 140 | 40 | 100 | Wed |
House blend beans | 9kg | 78kg | 82kg | 25kg | 60kg | Tue |
Takeaway cups | 180 | 1,600 | 1,700 | 900 | 800 | Mon |
Oat biscuits | 6 | 95 | ? | 30 | ? | ? |
The last row is the useful one. The model was told not to guess, so it did not, and now you know there is a disagreement in the oat biscuit figures that needs your judgement rather than a calculation. An output with no question marks in it, drawn from imperfect data, is a warning rather than a success.
Three things to check before you order
**Check one calculation by hand.** Pick a row, redo the arithmetic yourself. If it is right, trust the rest. If it is wrong, the error is systematic and you have saved yourself a bad order.
**Check the direction of the adjustment.** A busy week should raise numbers. If anything went down, something in the input was misread, usually a column the model interpreted as a total rather than a daily figure.
**Check the lead times against reality, not the contract.** Suppliers quote lead times they hit in a normal week. A busy week for you is often a busy week for them. If the deadline is Wednesday, order Monday.
What this does not handle
Be clear about the limits, because stock planning punishes overconfidence:
**Perishables.** Over-ordering fresh stock converts a cash problem into a waste problem. Order conservatively on anything with a short shelf life and accept selling out.
**New items with no history.** There is nothing to extrapolate from. Order small.
**Genuinely unusual events.** A one-off nearby event with no precedent is a judgement call. The data cannot tell you about something that has never happened.
**Substitution.** If oat milk runs out, some customers buy soy instead. No model knows your customers' substitution behaviour, and it changes the numbers meaningfully.
Once the order list is settled, the same stock facts can drive the board: planning a weekly specials menu with AI shows how to turn items near their use-by date into specials.
Make it repeatable
Save the prompt. Next quarter, change the dates and the event, paste fresh figures, run it again. The value compounds because your sales history gets longer and the comparisons get better, and because you stop rebuilding the same spreadsheet each time.
Keep the output. Next time, compare what you ordered against what you actually sold. Two or three rounds of that and you will know whether to adjust the model's numbers up or down, which is a thing only your own records can tell you.
Planning a seasonal promotion is the demand-side companion to this, and reading receipts and invoices covers getting supplier figures into a usable shape in the first place. If you are weighing what any of this should cost, how much a small business should spend on AI tools sets a sensible ceiling, and the small business guide covers what else is worth automating.
Sources: for the broader statistical caution about extrapolating from short histories, the UK Office for National Statistics publishes guidance on interpreting short-run trends that applies equally to a shop's own figures.
How did this land?
About the author
Developer Advocate
Steve builds something with Swarmz every week and writes up what worked, what broke, and what he'd do differently. Tutorials and hands-on guides are his lane.
