Part 2 of 3, distilled from our May 2026 online event AI in Demand Planning. A 30-minute conversation between Nicolai Harsbo Pedersen (Client Solution Manager at Hakio, 8 years in fashion operations) and Glen Roth (Enterprise Sales Director). This post focuses on the second of the three things they covered: how AI helps fashion brands catch overproduction risk earlier, by surfacing the signals most forecasts compress into a black box. 

Watch the full replay here.

The black box problem

Most fashion forecasting is built in spreadsheets, and most spreadsheets do the same thing with demand signals. They take many inputs, compress them into one number, and present it as the forecast. The signals disappear inside the model.

That's the black box, and Nicolai is direct about why it's a problem.

"You try to mirror as many compositions as you can and then condense them into one forecast, and that becomes sort of a black box. Very hard to iterate on. Very hard to correct at the end of the day."

When the forecast is wrong, and forecasts in fashion are often wrong, there's no way to look at the model and ask which signal misled it. The number is just the number. The planner either accepts it or rebuilds it from scratch.

That dynamic is part of why some brands have been cautious about adopting AI. If you don't trust your current black box, replacing it with a more sophisticated black box doesn't fix anything. It just makes the box harder to argue with.

What signal decomposition actually means

This is where Nicolai's framing of AI shifts. The point isn't a smarter forecast number. The point is to portray the composition of that number back to the planner.

"Rather than treating it as a black box, AI can portray to you the composition. Decompose this entire build of a forecast."

In practice: a forecast isn't one number, it's the sum of many components stacked on top of each other. Some of those components are obvious, like baseline demand and seasonality. Some of them are signals planners have always known about but never been able to see, separately, in the model.

The conversation surfaced two of those signals specifically.

Two signals most forecasts swallow silently

Discount-driven sales

The first is the difference between what sold organically and what sold because it was discounted. Manual processes typically don't separate them. The model sees demand. The reality is demand-at-a-price-point.

That distinction matters because it changes the buy.

"If you continuously buy into a product (or a similar product) that is being sold at a discount, you will end up either overstocking a product or you will really hurt your margin."

The black box hides which products are doing well and which products are only doing well on promotion. Decomposing the signal makes that visible: every time a forecast is presented, the planner can see how much of last year's number was healthy demand and how much was discount-pulled volume.

Stockout imputation

The second signal is the inverse problem. If you ran out of a product, you sold less than you would have. The historical sales number understates true demand. Manual processes can't really quantify the gap, so they tend to ignore it, which means the next buy is anchored to a number that's already too low.

"If you experience a lot of stockouts in your sales, on important product, you would want to hedge against that and make sure you have the appropriate availability in the future. That's where stockout imputation becomes really important, and that AI can do really brilliantly."

Stockout imputation is the technique of estimating what would have sold if the product had been on the shelf. It's not the kind of work that scales by hand. AI can do it, and the result is that the next buy is anchored to true demand rather than to suppressed sales data.

From intuition to evidence, without losing the gut feel

The risk in any "data-driven" pitch is that it sounds like AI is meant to overrule the experienced buyer. Nicolai is explicit that it shouldn't.

"It should support the intuition of an experienced buyer. It should not replace it. When you've had 10 years, 20 years in a category, you have a knowledge AI doesn't have. You have gut feeling, you have intuition, which is still very, very important for this."

What changes is the starting point. Instead of a planner opening a blank spreadsheet and rebuilding the forecast from scratch every season, AI puts a recommended forecast on the screen, with its components visible. The planner's job becomes refinement, not construction.

"React and refine, rather than construct."

It's the same line that ran through Section 1. In this section it has a specific application: you're not refining a single number, you're refining a decomposition. The components are visible, so the planner can challenge specific assumptions, accept others, and override the model where their judgement says to. All in the same view.

Trust is what unlocks the rest

The thread that runs through the whole section is trust. As Glen framed it, opening the black box is what allows teams to actually trust the AI, because they can see what's inside it.

"When you can show your team what is actually driving the numbers, they can trust it, challenge it, and adjust it intelligently."

That last word is the load-bearing one. If a team can't see the components of a forecast, "challenge" is hard and "adjust" is even harder. Decomposition turns a yes-or-no relationship with the model into a working dialogue. The planner contributes the things AI doesn't know: the buyer's instinct on a category, the commercial context behind a discount strategy. The model contributes the things humans can't track at scale.

Three things to take with you

1. Signal decomposition is what makes a forecast trustworthy. A forecast that's just a number is hard to challenge. A forecast that shows its components can be argued with, refined, and ultimately trusted.
2. AI should anchor planner judgement, not replace it. Twenty years of category experience is knowledge AI doesn't have. The point of AI is to do the work humans can't do at scale, not the work they're already good at.
3. Explicit recommendations make deviations traceable. When a planner overrides the model, the override is visible, not buried in a spreadsheet.

This is the second of three posts in the series. To hear the full 30-minute conversation in Nicolai and Glen's own words, rewatch the online event here.