What is Causal AI ?

Causal AI grounds decisions in causation rather than correlation to pinpoint what works and why. It enables
high-precision decision-making that accelerates research and development, promotes revenue growth, and
maximizes return on investment.

【 Use Cases by Field 】
• Manufacturing: Identify the causal drivers of yield and defects, and derive optimal process conditions.
• Marketing: Measure the true effect of ads and campaigns and design segment-specific strategies.
• New business & service design: Validate the causal impact of concepts and prototypes to raise the odds of
success.
• Healthcare: Estimate treatment effects with confounders adjusted, then tailor care by patient profiles.
• Finance: Map causal links among behaviors and market factors to reduce risk while improving profitability.
• Public policy: Evaluate programs in education, welfare, and environment with scientific causal evidence.

Causal AI Overview

Gartner describes Causal AI as follows:
"Causal artificial intelligence (AI) identifies and utilizes cause-and-effect relationships to go beyond
correlation-based predictive models and toward AI systems that can prescribe actions more effectively and
act more autonomously. It includes different techniques, such as causal graphs and simulation, that help
uncover causal relationships to improve decision making."

Interpreted for practitioners, Causal AI:
• Goes beyond correlation-based prediction - from "what will happen" to "why it happens."
• Identifies and exploits causal structure - via causal graphs and simulation̶to estimate the effect of actions.
• Elevates decision quality - by prescribing "what to do next," not just forecasting.
• Increases autonomy - systems that understand causes can choose and adapt actions more effectively.

Why causality: correlation doesn't explain "why"

Traditional statistics and machine learning primarily model correlation̶variables moving together. For
example, "regions with ads saw higher sales" could reflect ad impact or simply richer markets. Causality
asks, "If we intervene̶run ads̶how does the outcome change?" Answering that requires modeling and
adjusting for important covariates and confounders.

The difficulty - and the "fundamental problem of causal inference"

To measure causal effect rigorously, we would compare the same unit under treatment and no treatment
under identical conditions. But we cannot observe both worlds for the same unit at the same time̶this is the
fundamental problem of causal inference.
Randomized Controlled Trials (RCTs) were developed to address this problem at the group level by randomly
assigning treatment/control while averaging away background differences. Yet RCTs are often infeasible due
to ethical concerns, specialized infrastructure, cost, or time. Hence the field of causal inference, which
advances methods for estimating causal effects from observational data, adjusting for differences captured
in the data.

Covariates and Confounders

• Covariates: Background factors related to the variables of interest (age, gender, income, region, seasonality,
etc.).
• Confounders: Variables that influence both the cause XXX and the outcome YYY. Leaving them unadjusted
introduces bias into the estimated causal effects.
Causal inference must include and adjust key covariates and confounders in the model to recover credible
causal effects.

Estimating causal effects: Adjustment formulas

Structural causal inference identifies intervention effects by adjustment formulas that control confounders. A
fundamental example is the back-door criterion:
Adjustment formula :
P(Y | do(X = x)) = Σ_z P(Y | X = x, Z = z) · P(Z = z)
Here, do(X = x) denotes an intervention and Z is a set of confounders. We compare the effect of X within each
Z = z stratum and average by the population distribution of Z. Beyond back-door, the front-door criterion and
other identification strategies can recover causal effects when some confounders are unobserved or
structures are more complex.

Causal discovery and causal inference - better together

・Causal Discovery ：
Causal discovery: Algorithms that infer causal structure (arrow directions) from data. Useful for exploring
unfamiliar domains, but they depend on statistical assumptions and algorithmic constraints and may not
always yield correct results. Treat outputs as hypotheses and starting points.
・Causal Inference ：
Given a (theorized or vetted) causal structure, estimates effects by adjusting for covariates/confounders
and answering "What changes if we act?"

Causal discovery and causal inference are complementary. In practice, an effective cycle involves setting
rules reflecting domain knowledge, obtaining candidate structures through discovery, estimating effects
through inference, and incorporating the results into subsequent exploration and design.

Combining Structural Causal Models (SCM) with machine learning

SCMs make causal structure explicit, enabling reasoning about interventions and counterfactuals beyond
correlation. Coupled with modern ML, they scale to nonlinear relationships, high-dimensional covariates, and
large datasets. The following are examples of features.

[ Conditional Average Treatment Effect (CATE) ]
Beyond overall ATE, CATE captures segment-specific effects, e.g., differences by age group, income band, or
process temperature. CATE reveals who benefits from what, powering targeted marketing and process
optimization.

[ Root Cause Analysis (RCA) ]
RCA augmented with causal inference isolates actionable root causes, not just correlates, focusing limited
resources on the interventions that truly move outcomes.

How this differs from SEM(Covariance Structure Analysis) and Bayesian Networks

・Structural Equation Modeling (SEM) or Covariance Structure Analysis :
Primarily fits correlation structures; not designed for interventions or counterfactuals.
・Bayesian Networks：
Encode conditional dependencies; without causal commitments they cannot identify do(·) effects.

Causal AI, in contrast, makes causal structure explicit and estimates intervention effects by design.

Summary

Causal AI distinguishes correlation from causation, identifies effects via adjustment formulas (e.g., back-door
and front-door), properly adjusts confounders and covariates, combines causal discovery with causal
inference to reveal unknown structures and effects, and integrates SCM with machine learning to estimate
both ATE and CATE and to power RCA̶ultimately elevating decision-making across business and operations.