For as long as most of us have worked in business, demand forecasting has lived inside one department: supply chain. Ops guessed how much to order, how much to stock, how much to make. Finance found out the number after the fact — usually when it showed up as an unpleasant working-capital surprise. HR found out later still, scrambling to staff up or down once the demand shift had already landed on the shop floor.
That sequencing — forecast, then react, then react again, one department at a time — is what AI is quietly dismantling. Not primarily because the forecasts are more accurate now, though they usually are. It’s because the same forecast can now feed three or four decisions simultaneously, instead of triggering three or four separate scrambles in sequence.
The old model was a relay race. The new one is a shared instrument panel.
Picture a fairly ordinary planning meeting. Sales flags a demand spike coming in six weeks. By the time that signal works its way through the supply chain’s reorder logic, into finance’s cash-flow model, and finally into HR’s shift-planning spreadsheet, weeks have passed — and each department has quietly made its own slightly different assumption about how big the spike really is. Nobody’s wrong, exactly. They’re just not looking at the same number at the same moment.
A real case worth studying.
Danone, the French food and beverage group, ran into a version of this problem, and it’s worth revisiting because it shows the pattern clearly. According to a case study compiled by Best Practice AI, a large share of Danone’s volume moves through short shelf-life, promotion-heavy products — discounts, seasonal pushes, media-driven spikes — which made traditional forecasting methods unreliable and largely ad hoc. Danone’s response was a machine-learning forecasting system, and the detail that matters most for executives isn’t the model architecture — it’s what happened downstream. The same forecast began feeding sales, supply chain, finance, and marketing planning together, rather than each function adjusting its own separately derived number after the fact.
What changes for finance:
When the forecast is shared rather than relayed, finance stops being the last to know. Liquidity and working-capital planning can move from reactive to anticipatory — modelling cash needs against the same demand signal supply chain is already acting on, instead of waiting for a purchase-order spike to show up in the numbers weeks later.
What changes for HR and workforce planning:
The same logic applies to headcount and shift planning. Instead of scaling labour up or down after a demand shift has already strained (or idled) the floor, workforce plans can move on the same timeline as the forecast itself. This matters more than it sounds — labour decisions made under pressure tend to be expensive ones, whether that’s overtime premiums or last-minute contract staffing.
What changes for supply chain:
Supply chain’s role doesn’t shrink in this model — if anything, it becomes the natural owner of the shared signal, since demand sensing sits closest to the operational data. What changes is that supply chain stops being the sole translator between “what customers are about to want” and “what the rest of the business should do about it.”
The catch, because there’s always a catch.
None of this works if the forecast is wrong and nobody notices for a quarter. A shared bad number is worse than three separately-adjusted mediocre ones, because it propagates faster and gets challenged less — everyone assumes someone else already checked it. If you’re considering this for your own organisation, the governance question matters as much as the modelling question: who owns the forecast, who’s accountable when it drifts, and how quickly does a correction actually reach every function using it.
There’s also a people question underneath the technology one. A shared forecast only creates value if finance, HR, and supply chain actually trust the same number enough to act on it together — which usually means someone has to build that cross-functional habit deliberately. The software doesn’t do that part.
Where I’d start, if I were running this.
Not with a full enterprise rollout. Pick one high-volume, high-impact product line or category where demand swings are frequent and expensive to get wrong. Get supply chain, finance, and HR looking at the same forecast for that one category before you expand it anywhere else. The value isn’t in the algorithm being clever — it’s in three departments finally reacting to reality on the same clock instead of three different ones.
The compass, as always, isn’t the destination. It just tells you which way is actually forward — and in this case, forward means fewer departments finding out last.
Have you seen forecasting used this way inside your own organisation — as a shared signal rather than a supply chain tool? I’d be curious to hear how it played out.
Source: Danone case study details on AI-driven demand forecasting, from Best Practice AI: https://www.bestpractice.ai/ai-case-study-best-practice/danone_reduces_forecast_error_and_lost_sales_by_20_and_30_percent_respectively_and_achieves_a_10_point_roi_improvement_in_promotions_with_machine_learning
