Would you order stock from an AI sales forecast?
Suppose a café asks an AI assistant how many pastries to order next week. A neat daily estimate is easy to put into the supplier form. It doesn't tell the owner how much food could be left over, or how early they might sell out. Forecasting: Principles and Practice explains why forecasts need uncertainty alongside the central estimate. A prediction interval describes a range of possible future values at a stated probability, under the model's assumptions. Asking a chatbot to add a plausible-looking high and low number doesn't establish such an interval. Before changing the order, I'd ask where the estimate and range came from, then compare forecasts made before several past weeks with what actually sold. Keep later sales hidden from the forecasting step. The book makes that distinction too: fitting old data well is not evidence of forecasting unseen data well. For this imagined café, I'd put leftover food and days that sold out beside forecast error. A sold-out day hides demand the till couldn't record. Ordering too much and too little have different costs; the central estimate alone cannot choose between them. That's a purchasing check I'd propose, not a test of a particular AI tool. Would the forecast change your order, or just give your existing guess a more convincing presentation?
Comments
I'd put the supplier's delivery days next to that forecast. Suppose the café buys frozen pastries by the case, with deliveries only on Mondays and Thursdays. A daily sales estimate still leaves someone to work out which delivery covers the weekend and how many whole cases to buy. I'd want the assistant's proposed order to name the delivery date, the days it needs to cover and the case size, before the owner approves it. That's a hypothetical ordering check, not something tested here. Otherwise a sensible prediction can become an unusable order when it reaches the supplier.
I'd put a very ordinary rival through that same ordering exercise: use last week's sales for each matching weekday. The forecasting book calls this a seasonal naïve method—repeating the last observation from the same seasonal period—and recommends simple methods as benchmarks. That doesn't make last Tuesday a reliable guide to every Tuesday. It gives the AI forecast something cheap to beat. For this imagined café, compare both over the same unseen weeks, with the same delivery days and case sizes, then look at leftovers and missed sales. If the more elaborate forecast doesn't improve the buying decision, the owner has a reason to keep the simpler rule. That's a comparison to try, not a result established for this café. Better-looking numbers shouldn't be enough to add another subscription to a small shop's bills.
If the supplier offers the forecasting assistant, ask whose costs it counts. In this imagined café, a bigger order earns the supplier more while the café pays for anything left unsold. That doesn't make the forecast wrong; it makes ‘recommended order’ worth questioning separately. I'd give the comparison a cash limit as well as delivery constraints, and keep the owner's waste costs in the decision. A supplier's tool could still help. It shouldn't get to decide how much of the café's money is worth risking.
I'd include the time spent feeding each method in that comparison. Suppose the AI forecast needs someone to export till sales, rename products and explain promotions every week, while the simple rule uses a report the café already has. Have the person who places the order do both during the trial, and count that preparation alongside waste and missed sales. The extra work might be worth it if the orders improve enough. But I'd want to know before buying whether we're removing a weekly chore or giving the closer another spreadsheet to finish.