The bottom line
Predictive analytics answers 'what will sales be'; causal AI answers 'what drove them, and what would have happened otherwise'. For FMCG that distinction is the whole game, because promotion, distribution and price move together and a correlation cannot tell you which did the work. Causal methods — proper baselines and counterfactuals, controlling for confounders like seasonality and competitor activity — estimate incremental lift: the sales that would not have happened without each lever. That turns 'the promo drove 12%' into 'the promo drove 4 points, new distribution 6, and a competitor stock-out 2', which changes where the next pound goes. The honest limit: causal inference needs good data and defensible assumptions, it is not a black box, and it estimates effects with uncertainty rather than proving them.
In This Article
Three teams, one correlation, three stories
Every FMCG business has the same monthly meeting. Sales are up, and each function has a reason that happens to credit its own work. None is lying — they are all looking at a genuine correlation between their activity and the sales line. The problem is that promotion, distribution and price almost always move at the same time, so correlation cannot adjudicate between them, and the loudest or most senior voice usually wins the budget.
This is not a reporting gap that a better dashboard fixes. It is a causal question wearing a BI costume: of the sales we saw, how many would have happened anyway, and how much did each lever actually add? Answering it requires estimating what did not happen — the counterfactual — which is precisely what causal methods are built for and what predictive models are not.
Predictive and causal answer different questions
A predictive model is optimised to forecast an outcome accurately — next month's volume, next quarter's demand. It can be excellent at that while being useless for decisions, because a variable that predicts well is not necessarily a lever you can pull. Ice-cream sales predict sunscreen sales; neither causes the other, and no amount of predictive accuracy tells you that discounting ice cream will not move sunscreen.
Causal AI is optimised for a different thing: estimating the effect of an intervention. It asks what happens to sales if you run the promotion versus if you do not, holding everything else constant. For FMCG planning — where the whole point is deciding which lever to pull next — the causal question is the one that matters, and using a predictive model to answer it is a category error that produces confident, well-fitted, wrong conclusions.
Predictive asks 'what will happen'. Causal asks 'what happens if we do X instead of Y'. FMCG spends its money on interventions, so it needs the second question answered — a great forecast of the wrong question is still the wrong answer.
Incremental lift and the baseline that defines it
The number FMCG actually needs is incremental lift: the sales caused by an intervention over and above what would have happened without it. That requires a credible baseline — a counterfactual estimate of the no-promo world — because lift is the gap between what happened and that baseline, and a wrong baseline produces a wrong lift however sophisticated the model on top.
A naive baseline (last month, or the same month last year) ignores everything else that changed and systematically mis-attributes. Causal methods build the counterfactual properly, using comparable un-promoted products, stores or periods as controls, so the baseline reflects what the promoted item would have done under normal conditions. The output is defensible: 'this promotion drove 4 points of incremental volume; the headline 12% also included 6 points of new distribution and 2 from a competitor stock-out'.
Controlling for the things that move together
The reason the monthly argument is unresolvable is confounding — variables that influence both the lever and the outcome, so their effects are entangled. Seasonality, competitor promotions and stock-outs, weather, price changes and distribution shifts all move with sales and with each other. A method that does not explicitly account for them will hand the promotion credit that belonged to the weather or to a rival's empty shelf.
Causal AI's job is to disentangle these — to estimate each lever's effect while holding the others constant, using the structure of the data and defensible assumptions about what causes what. This is where the discipline lives: the model is only as good as the confounders it accounts for, which is why it needs rich data (distribution, pricing, competitor signals, not just your own sales) and a practitioner who understands the category, not just the algorithm.
What it changes in practice
When lift is decomposed honestly, budgets move. A promotion that looked like a 12% winner but drove only 4 points of genuine incremental volume — the rest being distribution and a one-off competitor gap — may not be worth repeating at that depth, while the distribution gain that quietly did half the work deserves more investment. Trade-spend effectiveness, the perennial FMCG black hole, becomes measurable rather than asserted.
It also changes forecasting: a demand forecast built on causal drivers is more durable than one that has absorbed a spurious correlation, because it understands which relationships will hold when conditions change and which were artefacts. On the Microsoft stack this sits naturally in a Fabric lakehouse that already unifies sales, distribution and pricing data, with the causal models running against that governed foundation rather than against a spreadsheet of one team's numbers.
Where this breaks
Causal inference is not a black box you point at data. It rests on assumptions about what causes what, and those assumptions have to be defensible and stated — a causal estimate presented without its assumptions is as untrustworthy as the correlation it replaced. It needs a practitioner who will be honest about the uncertainty rather than selling a single confident number.
It is also data-hungry in a specific way: you cannot separate promotion from competitor activity if you have no competitor data, and you cannot control for distribution if distribution is not measured cleanly. Many FMCG businesses discover the real gap is the data foundation, not the modelling. And causal estimates carry uncertainty ranges — they improve decisions by narrowing the argument and quantifying it, but they estimate effects, they do not prove them, and anyone claiming certainty is overselling.
What to do first
Start with the data that makes causal attribution possible, and one well-bounded question.
- Pick one recurring argument — usually promotion effectiveness — and define the lift question precisely
- Check you have the confounders: distribution, pricing, and some competitor signal, not just your own sales
- Unify those sources into one governed foundation (a Fabric lakehouse) so the model runs on trusted data
- Estimate incremental lift against a proper counterfactual baseline, with the assumptions stated
- Review the decomposition with the three teams who were arguing — the point is a shared, defensible number, not a black-box verdict
If promotion, distribution and pricing argue over the same sales line every month, the answer is causal, not another dashboard. Book 30 minutes with Amit on separating real lift from correlation across your FMCG data. No slides, no pitch deck, no obligation to proceed.
Free Assessment
Where does your operation sit on the data maturity curve?
8 questions. 3 minutes. You get a scored breakdown across data infrastructure, analytics readiness, and automation potential — with a specific next step for your industry.