Causal AI: What It Is and What It Buys a Business Decision
Three price points sit on a CPG pricing lead's desk before a shelf reset; none of last year's sales data covers any of them. Causal AI answers what happens if we act, where predictive analytics answers what usually happens. What causal AI buys that lead is an estimated effect for the action on the table, with a confidence interval, derived from a randomized experiment rather than a correlation in historical data.
- The causal claim comes from randomizing which price, claim, or pack a respondent sees inside an experiment. McFadden discrete choice, Mixed Logit, and ICLV are the estimators that read the result afterward; the randomization is what makes the read causal.
- On one published study, the best configuration reaches 0.832 rank correlation against a human baseline of 0.959, 87% of that measured ceiling; across the 43 studies passing design filters, the mean is 0.73 (causal fidelity paper).
- Causal AI is not causal discovery from observational data alone, and a validated replication score is not a guarantee for a market that hasn't been tested.
- The concrete case in this article: a pricing lead choosing among three shelf prices, walked through Judea Pearl's three rungs of causal knowledge - seeing, doing, imagining.
What is causal AI, and how is it different from predictive AI?
Causal AI estimates the effect of an action a business has not yet taken, using a randomized experiment rather than historical correlation. Predictive AI - sales forecasting, marketing mix models, most machine learning classifiers - describes what has usually happened given the conditions on record. Judea Pearl and Dana Mackenzie organized this distinction into three rungs of a ladder: seeing (association), doing (intervention), and imagining (counterfactual) (The Book of Why, 2018). Predictive AI lives on the first rung; causal AI operates on the second, by randomizing the action inside an experiment, and can extend to the third, by estimating what an action not taken would have produced. The tool category matters less than which rung the method actually reaches.
The decision: three prices, one shelf reset
The pricing lead's problem has a fixed shape: three candidate prices, a shelf reset date, and no history at any of the three price points. Historical volume data exists for the prices that were actually charged in the past, confounded by promotions, seasonality, and competitor moves at the time. None of it says what would happen at a price never charged. The question that matters is which of these three prices, if charged, maximizes volume and margin together. That is a rung-two question, and it requires a rung-two method.
Rung one: what historical sales data and MMM can see
Historical sales analytics and marketing mix models are built for association: they describe which past prices correlated with which past volumes, under the conditions that happened to hold at the time. That is useful for monitoring a trend or forecasting near-term demand inside the range of prices already charged. It cannot choose among three untested prices, because extrapolating a correlation outside the range it was measured in mistakes an association for an effect. A price move confounded with a promotion, a competitor stockout, or a seasonal shift produces a correlation that has nothing to do with the price itself, and rung one cannot separate the two.
Rung two: where causal AI operates
This is where causal AI operates: randomize which of the three prices a respondent sees inside a simulated market, then estimate the effect of price on choice with a confidence interval. McFadden discrete choice, Mixed Logit, and ICLV are the estimators that produce that effect from the experiment's data; none of them is a causal method on its own. The causal claim comes from the randomization in the experiment design - price assignment did not depend on anything else about the respondent or the market - and the estimator reads the effect out of data that randomization already made interpretable.
Preference-share and substitution questions between the three prices raise one more issue. A flat logit model carries the independence-of-irrelevant-alternatives assumption: it can misstate substitution patterns when the alternatives are not equally distinct from each other. Mixed Logit and ICLV relax that assumption, at the cost of more parameters to estimate.
The confidence interval that comes out of this experiment covers the effect within the simulated population tested, not the real market unconditionally.
Rung three: the counterfactual - what the price you didn't pick would have done
Rung three asks what would have happened under an action not taken, and a well-designed randomized experiment answers it for all three prices at once, not just the one eventually chosen. Because each arm was assigned independently, the same dataset that identifies the effect of the chosen price also estimates what the other two would have done, which is what the next reset needs. Advertising measurement does a version of this after the money is spent: GeoLift builds a geo-level counterfactual by weighting untreated regions to approximate what the treated region would have done without the spend (GeoLift methodology). Running the experiment before the reset gets that same comparison for all three prices while the budget is still uncommitted.
How does causal AI compare to predictive AI and causal-inference libraries like DoWhy and EconML?
Causal AI, predictive AI, and causal-inference libraries answer different parts of the same problem, and picking the wrong one for a given decision produces a confident answer to the wrong question.
| What it answers | Where the causal claim comes from | Best for | |
|---|---|---|---|
| Predictive AI (MMM, sales forecasting, ML models) | What usually happens, given the conditions on record | Correlation in historical data; no randomization | Monitoring trends and forecasting demand inside the range of prices already charged. |
| Causal-inference libraries (DoWhy, EconML) | The effect implied by a specified causal graph, often applied to observational data | The graph's assumptions, which are not directly testable from the data alone | Teams that already have (or are willing to assume) a causal graph and need effect-estimation code on data they already hold. |
| Causal AI (randomized experiment platforms) | What happens if a specific, untested action is taken | The randomized manipulation inside the experiment design | A decision-maker choosing among specific untested actions - a price, a claim, a pack - before committing budget. |
What causal AI is not
Causal AI is not causal discovery from observational data alone. Causal discovery libraries infer a plausible graph structure from patterns in existing data, and that graph is a hypothesis, not a randomized test.
Causal AI is also not the estimator. McFadden discrete choice, Mixed Logit, and ICLV are statistical tools for reading an effect out of experimental data; none of them manufactures a causal claim by itself, and applying one to observational data does not make the result causal.
A related failure mode is trusting stated preference on its own. Asking people what they would pay runs high against what they actually pay, with a median hypothetical-to-actual ratio of 1.35 across 28 studies (Murphy et al., Environmental and Resource Economics, study), unless the design is incentive-aligned so the stated answer has a real consequence.
No validation score is a guarantee that a method proven on past studies will hold in a market it has never been tested against. That gap is why the validation record has to be published rather than asserted.
Why does the validation record matter more than the method name?
It matters because a method's name - discrete choice, Mixed Logit, ICLV - says nothing about whether its output tracks real human behavior; only a replication test against real humans answers that. Subconscious's best configuration reaches 0.832 rank correlation against a published human result on one study, where two independent samples of real humans reach 0.959 - 87% of that measured human ceiling. Across the 43 published randomized studies that pass its design filters, the mean rank correlation is 0.73 (causal fidelity paper), and every study behind that mean is published at the leaderboard, misses included. One caveat travels with these numbers: a published validation study can sit inside a model's training data, and a model scored on a study it has already seen is not a fair test. The protocol filters for that; the field has not solved it. More comparisons against this standard are in the methods and validation hub.
What should a pricing lead actually do before the shelf reset?
Check whether the method proposed for the decision has ever been randomized and scored against real human behavior, and ask to see the misses, not just the wins. A pricing lead can do this without engaging a vendor: list the three prices, define the outcome metric before running anything, and confirm the test assigns respondents to prices at random rather than reading last year's sales by tier. A method's score across dozens of prior studies predicts its reliability on a new decision better than its name does, so compare any proposal against the leaderboard. To talk through which price to test first, book time.