Causal AI: why prediction is not enough to decide promotions

Predicting how much you will sell is not the same as knowing what will happen if you change something. That difference separates a good forecast from a good decision.

• Forecasting demand does not tell you what happens if you change the price.
• Correlations confuse the promotion with what surrounds it.
• Causal AI estimates the effect against a scenario without the promotion, from your own data.

Predicting is not deciding

Many tools forecast demand: how much will sell next week if everything stays as it is. That helps with stock, but it does not answer a promotion's question: what happens if I cut the price by 15%? And if I run it on other dates?

That is a question of cause and effect. It means estimating how sales change when you move a lever, separated from everything else that changes at the same time.

The trouble with correlations

A model built only on correlations learns patterns such as 'when there is a promotion, more sells'. But promotions are not handed out at random: they run on strong dates, on products that already sell well or in response to a competitor. If the model ignores that, it credits the promotion with sales that would have come anyway.

Causal AI starts from that difficulty. It uses the history to build the scenario without the intervention, what would have happened without the promotion, and estimates the effect as the difference, accounting for season, trend, price and what happens across the rest of the catalogue.

From measuring to deciding

Once you can measure the effect of past promotions, you can simulate future ones: try depths and dates before launching and compare the net profit of each option. It is the step from knowing what worked to choosing what pays.

And because every estimate rests on real promotions from your store, you also know how much evidence stands behind each figure: a depth you have tried three times is not the same as one you have never run.

How Effecta applies it

Effecta trains its causal models on each store's history: sales by product and day, prices, promotions and costs. With them it measures every past promotion, simulates the next ones and proposes a weekly plan within your rules for margin, discount and stock, which your team reviews and approves.

Behind the engine is causal AI research from the Universitat de Barcelona, applied to business questions with real data.