A forecast that gives you one number is hiding the useful part.
"Can the AI just tell me how much to order?" It's the question we hear most from owners looking at demand forecasting. Here's roughly how the conversation goes, and why the answer we give is about ranges, sold-out days and who makes the call.
An AI demand forecast is most useful as a range, not a single number. Ask for a likely low and high for each product and period, and set an ordering rule per product group that weighs the cost of running out against the cost of leftover stock. Tell the forecast what it can't see, such as planned promotions, new stores, supplier delays and days you were sold out, which can make demand look lower than it was. Compare it against a simple baseline like the same week last year, review the misses that matter weekly, and keep a person accountable for the order.
"Can the AI just tell me how much to order?"
This is a composite of a conversation we've had, in one form or another, with several owners: a shop with a few locations, a wholesaler, a food business. Details are blended; the questions are real.
Owner: We spend hours every week working out orders. Someone told me AI can forecast demand now. Can it just tell me how much to order?
Us: It can help a lot. But "how much to order" is two questions stuck together. One is what you're likely to sell. The other is what you want to risk. A forecast can answer the first. The second is yours.
Owner: What's the difference, in practice?
Us: Say you sell a sandwich that only lasts a day, and a bag of coffee that lasts six months. Even if the forecast says the same thing for both, you'd order them differently. Run out of the coffee and you lose a sale. Overorder the sandwich and you throw it away. The forecast doesn't know which of those hurts more. You do.
"What does the number actually mean?"
Owner: The demos I've seen show one number. "Next week: 140."
Us: That's the part we'd push back on. A single number hides the most useful thing a forecast can tell you, which is how sure it is. "Probably between 110 and 170" is much more helpful than "140," because it tells you how much room to leave.
Owner: So I'd want the range.
Us: Yes. Ask any tool you look at for a likely low and high, not just a middle. And ask how it's worked out. If it can't give you one, or the range is suspiciously narrow every week, that's worth knowing.
Some products will have tight ranges: steady sellers that move the same amount every week. Others will be wide: new products, seasonal ones, anything that depends on the weather. That spread is information. It tells you where to be careful.
The forecast's job is to tell you how unsure to be. The order is still a decision.
Decision guide
One number vs. a range with a rule
| Consideration | "Next week: 140" | "Likely 110 to 170" plus a rule |
|---|---|---|
| Short-life product | Order 140, throw away the rest. | Order near the low end. |
| Product you can't run out of | Order 140, run out on Thursday. | Order near the high end. |
| New product | Same confident number. | Wide range; start cautious, review weekly. |
| When it's wrong | Nobody knows why. | The miss is outside the range, so it's reviewed. |
The forecast's job is to tell you how unsure to be. The order is still a decision.
"What doesn't it know?"
Owner: It's trained on my sales history, though. Isn't that enough?
Us: It's a good start. But your sales history has holes, and some of them are sneaky.
The biggest one is sold-out days. If you ran out of something by Thursday lunchtime last month, your records show low sales for Thursday afternoon and Friday. The real demand was higher. You just didn't have anything to sell. A forecast that learns from that history will learn to order less, and you'll run out again. It's worth marking those days so the model, or whoever builds it, knows sales were capped.
Owner: What else?
Us: Anything that's planned but not in the data yet. A promotion next month. A new location. A product you're dropping. A supplier who's told you deliveries will be late. Local events. A competitor closing down the street. The people who place orders know these things. The model only knows them if someone tells it.
So we'd always build in a way for the person ordering to say "this week is different, here's why" and see the forecast adjust, or adjust it themselves, with a note.
"So who decides?"
Owner: If I can override it, what's the point?
Us: The point is that you override it a few times a week instead of working out every line from scratch. Most products are boring, in a good way. The forecast handles those, and your time goes to the ones that need judgment.
What makes that work is a simple rule for each group of products, agreed in advance. For short-life items, order closer to the low end of the range. For things that keep, and where running out costs you customers, order nearer the high end. For anything new, start cautious and review weekly. Someone with authority writes those rules down, and the person ordering follows them unless they have a reason not to, which they note.
Owner: And if it's wrong?
Us: It will be, sometimes. That's why a person still owns the order. The forecast is advice with a range attached. It's the same question we wrote about in whether you need a model or a rule nobody wrote down: much of the value comes from writing down the rule.
"How do we know it's working?"
Owner: Vendors all say their accuracy is great.
Us: Ask what they're comparing against. The simplest test is to compare the forecast to a dumb baseline, like "same week last year" or "average of the last four weeks." If a sophisticated model can't beat that on your own products, it isn't earning its keep yet.
Then look at misses, not averages. An average accuracy figure can look excellent while the forecast keeps getting your three most important products wrong. Once a week, look at the items where the actual sales landed outside the range, especially the ones that matter most to you, and ask why. Sometimes it's a sold-out day. Sometimes it's something the forecast couldn't have known. Sometimes the model's wrong and needs fixing.
Keep a short log of overrides, too: what the forecast said, what you ordered, why, and what happened. After a couple of months, you'll know where you're better than the model and where it's better than you.
How the work moves
A weekly forecast review
- Mark sold-out days
So capped sales aren't read as low demand.
- Tell it what's different
Promotions, new locations, supplier delays.
- Order by the rule
Low, middle or high end of the range by product group.
- Review the misses
Items that landed outside the range, most important first.
- Check against a baseline
Same week last year, or the last four weeks.
"Where would we start?"
Us: With one category you know well and order often, and a few months of history where you've marked the sold-out days. Run the forecast alongside your normal ordering for a few weeks without acting on it. Compare both to what actually sold, and to the simple baseline. Then decide.
If it helps, our note on detecting inventory truth drift covers a common problem underneath all of this: whether the stock numbers the forecast starts from are right in the first place.
We build AI applications that show their uncertainty and leave the decision with the person responsible, and we connect them to stock and ordering in our fulfillment operations work. If you're weighing a forecasting tool, or wondering whether you need one, start the conversation here.
Related work: Howcast.
We build alongside the people doing the work.
Bring us the problem you’re thinking about.