Walk into almost any monthly planning review and you’ll see three common line items: expedited freight, markdowns to clear excess stock, and overtime to catch up on production. They are discussed as three separate problems. In my experience, they’re three bills for the same miss. Demand shifted underneath a number everyone agreed on, and nobody saw it until it showed up as cost.
The truth is, the forecast may have been accurate at a high level, but it didn’t explain what was driving demand, how those drivers were changing, or what decisions the business needed to make.
That’s why, in my work with supply chain teams, I advise leaders to step back from the usual debate about forecast accuracy. The bigger opportunity is to build on demand planning with demand modeling: understanding what’s driving demand, exploring the range of ways it could play out, and anticipating what could change before it shows up as cost. More importantly, those insights need to feed scenario planning, simulation, and optimization, so teams can evaluate real alternatives and choose the best one. That’s where the real business value gets unlocked.
When Right Still Goes Wrong
Here’s the version I keep seeing. Two regions are each forecast to sell ten thousand units each. Actual demand comes in at twelve thousand in one and eight thousand in the other. Add them up and the company-wide number is exactly right.
But a plan built to stock both regions equally creates a shortage in one and a surplus in the other. The business spends the next quarter paying for the gap through transfers, expedited freight, and markdowns. The forecast was accurate. It just didn’t carry the detail the decision needed. Of course, this is a simplified example. Now imagine that complexity multiplied across thousands of products, hundreds of locations, shifting demand patterns, and interconnected supply chain constraints. The impact compounds quickly, and so does the cost of getting those decisions wrong.
The same thing can hide inside a single growth figure. Ten percent growth can come from more customers, bigger orders, or existing customers ordering more often, and each calls for a different fulfillment footprint, delivery cadence, or inventory position. Collapse them into one percentage and you’ve erased the information a planner needed to act. More granularity shows where demand landed, but only the drivers behind it explain why.
From Planning the Number to Modeling the Demand
When most teams hear “demand,” they jump straight to the forecast. Demand modeling shifts the focus to the step before the number: what’s driving demand, in what context, and for which decision. It doesn’t replace demand planning, which is good at what it was built for, a shared number the business can run against. It adds a layer underneath that explains why the number looks the way it does and how it could move, with the planner’s judgment still central.
Some drivers are patterns already in the demand history, like seasonality. Others are levers the business controls, such as pricing and promotions, and others are forces it can’t control, like a jump in mortgage rates for a home goods company. Which ones matter depends on the decision: planning next week’s deliveries calls for a different view of demand than sizing a distribution center five years out. Get that context right, and a good forecast becomes the natural outcome rather than the whole exercise.
Seeing Change Before It Shows Up as Cost
Demand is also dynamic, and a number is only good until the assumptions underneath it change. I could tell you a product will sell 3,000 units next month at a given location, then realize two days later it’s headed for 10,000 because customer preferences shifted. Planning off a static number means finding that out in next month’s cost report.
The teams that get ahead of this watch the drivers, not just the number. Many external signals move weeks or months ahead of demand, giving early warning that a plan is drifting. When a driver moves, these teams don’t rebuild from scratch. They re-run scenarios they’ve already set up and see what changes while there’s still time to act.
Plan for a Range, Not a Point
No forecast will ever be perfect, and a single number implies a certainty the business doesn’t have. A range is closer to reality: a product might most likely sell 50 units a week but could reach 275 based on the underlying uncertainty in the signal. Knowing which way that range leans, toward upside or downside, tells you more than the midpoint.
That range is what makes demand scenarios useful. Teams can ask what happens if a promotion overperforms, interest rates move, or a key customer pulls back, and see which assumptions the plan is most sensitive to. The answers point to concrete moves: adding buffer, right-sizing inventory, or pulling production forward a few weeks.
It also changes how we should judge a forecast. You can improve accuracy by a point and still make the same decision you would have made anyway, so the improvement created no value. What matters is decision accuracy: whether the range led to a better call.
But a demand scenario only shows what demand might do. To know what the business should do about it, each scenario has to run through the real constraints of the supply chain.
Where the Value Actually Shows Up
I’ve seen this bullwhip effect across industries. Demand planning adds a little lift to protect service. Supply planning adds a buffer to the production schedule, and the plant runs a bit extra to be safe. By the time it reaches the warehouse, the business is overstocked on the wrong products in the wrong places, service still slips, and no one made a bad decision along the way.
That isn’t a forecasting failure. It’s what happens when demand, supply, inventory, and transportation decisions get made in separate rooms, each hedging against the others. Understanding demand is only half the equation; the other half is connecting it to the decisions it drives.
This is where simulation and optimization earn their place. An optimization model can take demand scenarios and find the best feasible plan across inventory, production, capacity, and transportation, within real limits on cost and service. A simulation can then stress-test that plan against the range, showing where it holds and where it breaks. Together, they turn a debate about whose number to trust into a comparison of real alternatives.
That’s the conversation executives want. What does it cost to hold service at 98 percent instead of 95? Is it cheaper to pre-build ahead of the peak or add overtime when it hits? Those answers come from decision models, but they’re only as good as the demand insight feeding them.
AI That Explains and Helps You Decide
AI now accounts for two-thirds of supply chain digital investment, yet Gartner found more than half of chief supply chain officers are unclear on the return. Much depends on what AI is asked to do.
Rather than forecasting each product and location in isolation, machine learning can model every level of the business at once, so patterns at one level sharpen the others: a red bike with six months of sales history can borrow from a blue bike with ten years of it. It can flag anomalies and tell a one-time spike from a genuine shift. And instead of a single number, it can produce a range of likely outcomes and carry that uncertainty from individual products up to category totals, where capacity and network decisions get made.
The bigger step is AI that helps decide, not just describe. An assistant that pulls up a forecast and explains the trend is useful, but explaining a number and evaluating a decision are different jobs. The stronger pattern takes a planner’s “what if,” runs it through simulation and optimization models against real constraints, and returns tested alternatives with their trade-offs. Both can arrive in the same confident sentence, so it’s worth knowing which one you’re getting.
People still belong in the loop. AI is good at finding correlations, but it takes domain judgment to tell whether a driver truly causes demand or just moves alongside it. Planners also need to see the evidence behind a recommendation, including what data it used and how current that data is. AI should make that judgment faster and better informed, not skip it.
Three Questions to Ask Before You Commit
Before I trust any demand-driven recommendation, I ask three things.
What’s driving this number, and what else could produce it? Different causes call for different actions.
Which alternatives were evaluated, and against which constraints? A single plan isn’t a decision.
What trade-offs matter, and what happens if the assumption is wrong? A good plan holds up across the range, not just at the midpoint.
What Accuracy Is Actually Worth
I reviewed an analysis for a Fortune 500 multi-brand retailer that estimated a six-point accuracy gain could be worth about $1 million a year per point, plus roughly $2 million in annual delivery savings. Those were modeled estimates, not booked results. The value didn’t come from accuracy alone; it came from weekly, territory-level demand feeding better delivery, labor, and vendor decisions.
Accuracy matters, but it’s an input. The test is whether a business understands what’s driving demand, plans for the range, and evaluates its options before committing. That’s where the value gets unlocked.
References linked in the text
- Lee, H. L., Padmanabhan, V., & Whang, S. (1997). Information Distortion in a Supply Chain: The Bullwhip Effect. Management Science. Stanford GSB (anchor: “bullwhip effect”)
- Gartner (August 5, 2026). Gartner Survey Finds Majority of Chief Supply Chain Officers Unclear on AI Investment Returns. Press release (anchor: “Gartner found”)
- Gneiting, T., & Katzfuss, M. (2014). Probabilistic Forecasting. Annual Review of Statistics and Its Application. DOI (anchor: “A range is closer to reality”)
- Hyndman, R. J., Ahmed, R. A., Athanasopoulos, G., & Shang, H. L. (2011). Optimal Combination Forecasts for Hierarchical Time Series. Computational Statistics and Data Analysis. Publication page (anchor: “model every level of the business at once”)
