Choreography: demand forecasting with your feet on the ground
AI builds demand scenarios fast, but the growth and seasonality assumption is yours. Auditing the forecast before you buy, scale, or hire is what separates a forecast from a confident guess.
Monday morning, and the operations director wants to know how many units you'll need to process during Black Friday month, to close the temp labor contract by Wednesday. You ask AI for a demand projection based on history, and in minutes a nice round number comes back, with a pretty chart and a confident sentence about "expected growth of 40% over the same period last year". The 40% seems reasonable, until you remember that last year had a stock rupture that dammed up orders, and it won't repeat. Hiring temp labor on top of that inflated number means paying idle people in December. Hiring too few means a queue and an unhappy customer on the date that matters most all year. The pretty forecast doesn't decide anything on its own, the assumption behind it is yours, and it's what you audit before you sign the contract.
You ask AI for the working capital projection needed for next quarter, based on sales history. It confidently returns a number assuming 15% monthly growth, but that number comes from an atypical month with a one off promotion that won't repeat. Auditing the assumption before committing cash prevents idle money sitting still or running short of breath at the wrong time.
To size the legal team for next semester, AI projects the volume of contracts to review based on the company's recent growth, assuming an expansion rate that actually came from a one off acquisition already completed. Hiring extra lawyers on top of that inflated number costs a payroll that doesn't hold up once the volume normalizes.
AI projects next quarter's lead volume to size the support team, assuming that a recent, one off campaign's conversion rate will repeat every month. Without auditing that assumption, the team hires people to handle a volume that only ever existed in that specific campaign.
To size how many openings to post next quarter, AI projects headcount growth based on recent expansion, without separating what was structural hiring from what was a one off backfill for a turnover spike already resolved. Hiring an extra recruiter on top of that inflated number costs budget for people who won't be needed.
To decide the infrastructure capacity for the next launch, AI projects the number of concurrent users assuming the same adoption rate as a recent viral feature, which isn't the product's normal pattern. Provisioning servers on top of that one off spike costs money every month, for the rest of the year.
For next quarter's commission forecast, AI projects sales volume assuming that an exceptionally good month's close rate will hold, without checking that that month had an industry trade show that won't repeat. The commission budget built on top of that number inflates the projected cost with no need to.
Thursday, distribution center capacity meeting, and you need to decide whether to open an extra shift next month. AI projects order volume based on the last six months and returns a base scenario of 12 thousand orders a day, with the confident line that it's "a direct extrapolation of the recent trend". The problem is those six months include a month when a big competitor went down for two days and pushed extra orders your way, a one off event that shouldn't become a trend. If you open the extra shift on top of that inflated number, you pay overtime for nothing all month. If you decide without auditing the assumption, you also risk underestimating and jamming the operation against a real demand that is, in fact, rising. The projection speeds up the math; knowing which piece of the history is noise and which is a real trend is still your job.
To size next year's internal audit team, AI projects the volume of controls to review assuming the same growth pace as an atypical quarter, which included the one off rollout of a new system. Auditing that assumption prevents oversizing a team for a volume that won't repeat.
To decide whether to scale the technical support team, AI projects ticket volume assuming a recent spike, caused by a bug already fixed, is the new normal. Hiring extra support on top of that inflated number costs a payroll that becomes excess as soon as the bug drops off the count.
To size how many usability testing sessions to schedule next quarter, AI projects research demand assuming the same pace as an intense launch period, which isn't the team's normal rhythm. Reserving research budget on top of that one off spike wastes resources that could go to another priority.
For the board to decide whether to enter a new market, AI projects the potential demand assuming an adoption rate that's "reasonable for the sector", without ever having seen the company's real conversion history in similar launches. Auditing that rate before the investment decision prevents betting on a market with a number nobody validated.
Monday morning, and the director wants the labor number for Black Friday closed by Wednesday. You ask AI for the projection, it comes back fast, with a pretty chart and a confident line about expected growth. The number looks ready to become a contract. The problem is that "looking ready" and "being right" are different things, and in a demand forecast the difference between the two costs people hired for nothing or an unhappy customer's line on the date that matters most all year.
The core idea of this lesson. AI builds a demand scenario fast: it structures the model, generates the variations, writes the explanation. That's a real gain. But the assumption behind the number, what growth is reasonable for your business right now, what in the history is a real trend and what was a one off event that won't repeat, stays yours. Auditing the forecast before you buy, scale, or hire is what separates a forecast from a guess wearing the look of certainty.
01The choreography: who does what
Building a forecast used to always be heavy work: pull history, adjust for seasonality, run the numbers for three scenarios, and still write the explanation for whoever decides on top of it. AI really changes that math. It structures the model, organizing the history into base, trend, and seasonality. It generates the variations, optimistic, base, and pessimistic, from an assumption you supply. And it writes the explanation, turning the number into text the director understands fast.
What stays yours, and gets more expensive precisely because the grunt work vanished, is the assumption. What growth is reasonable for your business, right now, not in general. What in the history is a real trend and what was a one off event, a stock rupture, an isolated campaign, a competitor going down, that shouldn't be counted on again. That's reading your operation, not calculation, and it's exactly what AI doesn't have when you don't hand it over.
02The scenario tree: optimistic, base, pessimistic
Demand isn't a number, it's three plausible paths from your audited history. The base scenario is your most honest bet for next period's volume, units to process, orders to deliver, tickets to handle. The optimistic one is the branch above, demand comes in stronger than expected. The pessimistic one is the branch below, demand slows down or a supplier fails.
The point of building all three isn't getting the future right, it's drawing the range within which your operation will likely need to respond. If even your pessimistic scenario leaves the operation standing, your capacity decision can be leaner. If the base scenario already pushes the operation to the limit, you decide with more of a safety margin. AI builds this entire tree in one request, but each branch depends on an assumption you need to have deliberately given it, otherwise it fills in a generic growth number that seems reasonable and isn't yours.
03Sensitivity: which variable moves the projected demand the most
Every demand forecast carries several variables, a big scheduled promotion, the possible rupture of a key supplier, the seasonality of a holiday date, the risk of losing a big customer. They don't weigh the same. Finding out which one really moves the result is what separates whoever decides with method from whoever just looks at the final number.
The choreography is direct: fix the base scenario and test one variable at a time. What if the scheduled promotion gets canceled, how much does the volume drop? What if the key supplier is two weeks late, is the impact big or small next to the total? What if the big customer cuts their order by 20%, does that change the total volume or is it a detail in the middle of the curve? AI recalculates each case fast. Your energy for reading the world should go to the variable that moves the result the most, the lever, not spread evenly across all of them.
04The core danger: AI invents demand with the look of certainty
Here lives the risk that gives this lesson its name. AI doesn't fail like a broken calculator. It fails with total confidence, handing over growth that's "reasonable for the sector" when your operation has a quite different pattern, or assuming a one off spike in the history is the new trend. It has no instinct to distrust its own number, because it never saw the stock rupture, the competitor going down, the isolated campaign that inflated last month.
In a demand forecast, that becomes an expensive decision. Hiring temp labor, buying stock, opening an extra shift, all of it on top of a number nobody audited is betting real money on an assumption that might not even be yours. And the decision is usually not very reversible: people hired and stock bought don't return easily when the spike doesn't come. The less reversible the decision the forecast feeds, the stronger your audit needs to be before you sign off on the number.
Do it now
Take a real capacity decision that depends on a demand forecast, your real task works well: hiring temps, opening an extra shift, buying stock, scaling support. Run the full choreography in four steps:
- YOUR ASSUMPTION: before opening AI, write by hand the growth rate or seasonality you consider reasonable for your operation right now, and one sentence on why that number is defensible in your case.
- TREE: ask AI to build the demand scenario with optimistic, base, and pessimistic using exactly your assumption, without letting it choose the growth on its own.
- SENSITIVITY: fix the base scenario and test, one at a time, the two or three variables that worry you most. Note which one moves the projected volume the most.
- AUDIT: look at the history that fed the model and flag any one off event (rupture, isolated campaign, competitor going down) that shouldn't become a trend. Only after that does the forecast get permission to become a capacity decision.
Practice
1. In AI powered demand forecasting, what stays yours and gets more expensive precisely because the grunt work of the calculation disappeared?
2. You tested the forecast's sensitivity and saw that a key supplier's delay moves the projected volume much more than a possible order cut from a big customer. What does that tell you?
For the board
On looking finishedlooking finished and being right are different things. In a forecast, the difference costs people hired.
On what is yoursthe AI cannot tell a spike caused by a stockout from a genuine trend. That reading is always yours.
On sensitivitythe variable that moves the volume most deserves the most defensible assumption and your attention first.
Thanks for the feedback. It helps sharpen the next lesson.