Causal Machine Learning: the new frontier of corporate decision-making

Discovering the cause-effect link: strategies to improve decision-making and make more accurate predictions

Published on Mit Sloan Management Review Italy, May/June/July 2025.

Themachine learningis now widely used forguide decisionsin processes where it is sufficient to measure the probability of a specific outcome, for example, whether a customer will repay a loan. However, because technology, in its traditional application, relies on correlations to make predictions, the insights it offers managers are imperfect at best when it comes to anticipating the impact of different choices on business outcomes (Feuerriegelet al., 2022a).

Consider the executives of a large company who must decide how much to invest in Research & Development (R&D) in the next year. Using the analysis methodMachine Learning (ML)traditional, they may wonder what will happen when they increase spending. They may find a strong correlation between higher levels of investment and higher revenues when the economy is growing. And they may conclude that since economic conditions are favorable, they should increase their R&D budget.

But should they really? And if so, by how much? External factors, such as consumer spending levels, technological spillovers from competitors, and interest rates also influence revenue growth. Comparing how different levels of investment might affect revenue, taking into account these other variables, is useful for the manager trying to determine the R&D budget that will bring the most benefit to the company.

The Causal ML, aemerging area of machine learning,can help answer these questionswhat ifthrough causal inference. Similar to how marketers use A/B testing to deduce which of two advertisements is likely to generate more sales,Causal ML can inform what might happen if managers were to take a particular action(Feuerriegelet al., 2022b).

This makes theuseful technology in manyof the samebusiness functions that use traditional ML,including product development, manufacturing, finance, HR and marketing (von Zahnet al., 2024). Traditional MLis stillthe ideal approachwhen the only objective is tomake predictions,such as whether stock prices will rise or which products customers are most likely to buy. When a company wantspredictwhat would happen if he made one decision over another, for example if a 10% or no discount is more likely to make a customer repeat purchase, he needs aCausal ML.

Our research into machine learning and AI and our experience helping companies apply causal ML point to apath to successfully use this technology(Box The research). Businesses will also need theright skillsand they will have toincrease employee causal ML literacy.

What causal ML can and cannot do

Causal ML is a powerful tool, but managers may find the name misleading. The label'counterfactual prediction'would more accurately reflect what it does:predict outcomes based on hypothetical actions.Technology is better understood as a way to make better hypotheses rather than as a source of definitive answers. By framing it this way, managers can be reminded todo not overinterpret the results.

It does this by usingcausal inference,which examines past results to understand cause-and-effect relationships between variables. Instead of focusing on why something happened, Causal ML applies these relationships to predict the effects of interventionsnew and future-oriented contexts.

However, the method cannot explain why a causal relationship exists between a particular factor and the outcome it affects. For example, a Causal ML model might predict that reducing an R&D budget would decrease revenue, but it would not explain why this relationship exists or whether confounding factors – which influence both the decision and the outcome – might change and invalidate the prediction. Managers shoulduse their experiencein the industry to evaluate whether a certain forecast makes sense. This approach helps ensure that model predictions are interpreted correctly and remain relevant to real-world decisions. Like traditional Machine Learning,Causal ML is more effective when managers have large volumes of data,options are clearly defined and the desired outcome is well understood.It is generally not suitable for decisionsone-offand scenarios that require intuition or creativity.

Choose the right problem and data

Causal ML is best suited topredict the outcomes of simple decisionswhich are supported by extensive historical data from internal and external sources. Operations questions may be good candidates for the approach because they are asked frequently and companies have a lot of data to support them (von Krogh, Ben-Menahem, & Shrestha, 2021). Below are some examples of using Causal ML in this context:

Booking.com collects data from thousands of hotel reservations every hour. The company's marketers use thecausal analysis methodto determine not only whethergrant discounts,but alsowhich customers should get them.Chocolate maker Lindt has extensive data on environmental conditions, equipment, packaging and other factors that influence the quality of its world-famous truffles. Production managers use theCausal MLto help them fine-tune parameters such as thetemperature of the machinesand themold configurationsfor truffles (ETH AI Center, 2023).Hitachi ABB Power Grids relied on Causal ML to reduce failure rates in its semiconductor manufacturing process, using machine performance data. It was able to halve the yield loss by identifying the combination of machines that consistently produced the best quality chips (Senoner, Netland, & Feuerriegel, 2022).

At Novartis, managers educated on the capabilities of different types of ML were able to identify several decision-making tasks where replacing traditional ML with Causal ML offered significant benefits. They had asked a traditional ML model whether increasing the marketing budget would increase sales, but its predictions didn't help them decide how to allocate that budget. They decided to use Causal ML to evaluate how different promotional campaigns could impact future sales. They used forecasts to allocate resources to the campaigns that were likely to be most effective.

A decision suitable for Causal ML can be expressed as a number or a binary choice(for example, an amount of revenue or a purchase/possession). It can also be phrased as a question about what action to take: allocate a marketing budget of $10,000 or $15,000 for the next quarter, or offer a 10% or no discount on a product (Wasserbacher and Spindler, 2022).

Also, the methodCausal ML cannot effectively address all potential use cases, even if apparently it seems suitable for this purpose.Confounders– the variables that influence both the outcome and the decision –introduce distortions that influence forecastsand must be taken into consideration. They can be difficult or impossible to test and affect the accuracy of predictions. If, for example, data is available only for product sales during an economic growth phase, forecasts of product sales during a downturn will be less reliable.

When managers have determined what they want to decide, identified how they will measure the outcome, and affirmed that they have sufficient data, they can begin working with data scientists to assemble and categorize the data to build their Causal ML model. Business leaders and other people with domain knowledge are essential partners of data scientists and ML experts in building causal ML models that deliver reliable results.

To train the model to graspcomplex cause-effect relationshipsdata from at least a few dozen, and ideally hundreds or thousands, of historical decisions are needed. With onemassive amount of data,the model can uncover connections between variables that may be unknown to managers or difficult to quantify. Less data leads to less accurate predictions.

In principle,the causal analysis method requires three categories of datawhich was previously mentioned:decisions, outcomes and confounding factors.Decision data includes what managers have done in the past, such as staffing levels or budgets set, discounts offered, investments made, or processes changed. Outcomes data can include any measurable business outcome, such as sales volume, revenue growth, quality metrics, or productivity.

Confounding factors can come from internal or external sources. They can include economic conditions, the composition of the workforce and the behavior of competitors, and can vary depending on the decision to be made. For a marketing decision, the type of device customers use can be a confounding factor, because those who own a more expensive smartphone may tend to spend more, regardless of whether they qualify for an incentive or not.

For example, the Neue Zürcher Zeitung, an international media company that publishes Switzerland's largest circulation newspaper, implemented Causal ML to improve the effectiveness of editors' content promotion decisions. The decision variable was the promotion of an online article on one of the two front pages that were served to readers. The outcome variable was a performance score that combined website traffic, reader engagement, and subscription subscriptions. Confounders included temporal factors (such as time of day), content characteristics (such as article format), past performance indicators (including clicks), and past promotion decisions (including whether the article had been promoted elsewhere).

Identify possible causal factors

A valuable lesson of our work was the usefulness of outlining acausal graphon a whiteboard illustrating the expected relationships between the outcome, the decision, and the confounding factors at the beginning of the model development process. Managers' knowledge and experience are essential here, because they have repeatedly made decisions and learned to anticipate certain outcomes.

The causal graph tells data scientists (who should be experts in causal inference) whethertreat a variable as a cause or effect in the model.This way, the team can rule out reverse causality errors. In other words, it can ensure that the model does not misinterpret one variable as causing another when, in reality, the effect is the opposite.

Imagine a celebrity with millions of followers on social media. If we don't know much about social media or celebrity, we might conclude that fame comes from having a high number of followers. The opposite is more likely to be true. As even the average teenager observed, to get millions of strangers to follow their social media accounts, they must first do something that gets them noticed. In the case of our question about R&D spending,the budget influences revenue, not the other way around.Meanwhile, confounding factors such as the economic climate, market trends or team expertise are recognized as factors that drive both the budget decision and business results, but are not influenced by either. The model would take all of this into account (Figure 1).

Choose the exit

Next, managers must choose thetype of responsethat the model must provide in response to the question (in statistics, the output or estimate): it can predict the final result of a decision or the relative benefit of an alternative compared to another.

Each of these results can be useful, depending on how the manager is thinking about a decision.Focus on the final results,such as potential revenues in different budget scenarios or incentives customized for individual customers, it helps in strategic planning. However, comparing the incremental effects of different decisions is often enough to make one: if a manager wants to know which of two advertisements can increase sales more effectively, he does not necessarily need to predict the amount of revenue that each variant could generate. He just needs to know the relative benefit: that one ad is capable of generating three times more revenue than the other. Furthermore,focus on the benefitsrelative generates more reliable predictions than focusing on the final results. We recommend pursuing only the necessary granularity.

The editors at the Neue Zürcher Zeitung were interested in predicting actual click-through rates for each promoted item, but the company chose instead to predict the likely net performance gain from promoting an item. This approach allowed Causal ML to make more accurate predictions about which content, if promoted, would increase clicks and subscriptions. Editors learned that promoting articles written by the editor-in-chief significantly increased both outcomes (Persson, Feuerriegel, & Kadar, 2023). The editors had promoted the editor-in-chief's articles sparingly, and the results served as a starting point for reviewing their promotional strategy.

Training, testing and validation of the model

Once managers havedefinedthe decision they want to makeand the type of output they prefer, data and machine learning scientists can choose theCausal ML model best suited to the job.Once the model is implemented, machine learning engineers will train it using previously categorized data.

The final phase is to test and validate the Causal ML model in practice, to ensure that it is reliable and that its predictions translate into better business performance. Validation also provides an opportunity for decision makers, including executives, to gain confidence in its predictions. Starting with relatively simple, linear problems, where clear decision alternatives can be identified and evaluated, makes this phase easier to accomplish.

Testing and validation require attentionbecause managers can only observe the outcome of the decision made in the real world. They have no way of knowing what the outcome would have been if a different decision had been made. Two strategies,human in the loopand the well-known A/B testing approach,have proven successful.

Neue Zürcher Zeitung has chosen to integrate the model's recommendations with human decision-making processes (Ibid). The Causal ML model recommends which content to promote, but editors make the final decisions. The model is based on the same information that editors previously used to make their promotion decisions; therefore, they can trust that the model is not missing key elements. Causal ML model recommendations typically match editors' feelings, which gives them confidence in the model's reliability.

Some decisions are difficult, and editors know their judgment is not perfect. In cases where Causal ML recommends a different decision than they would have made, editors can test the recommendation and see the result.Over time, they should see that the Causal ML method can provide reliable recommendations in ambiguous situations.Then they will be able to follow Causal ML's recommendations more often instead of their instincts.

Hitachi ABB used A/B testing to validate Causal ML models built to improve manufacturing quality. In one application, managers used the model to predict which of several machines would produce the best quality in the etching and implantation stages of the semiconductor manufacturing process, contributing to higher quality production overall. To confirm the reliability of the predictions, managers conducted a controlled experiment in which they changed the machine used for etching and implantation, while keeping the machines used for other processes unchanged. They found that the best machine for incision and implantation was the same one that the causal ML model had predicted. Thanks to Causal ML, managers were able to find and resolve the source of production problems more efficiently than they could with manual methods or traditional ML (Senoneret al., 2021).

Prepare the organization

Although Causal ML has the potential to improve decisions,implementing these systems requires a high level of AI literacy in the workforce,specialized technical skills and patience, becausethese projects can take longer to develop than traditional ML applications.Managers can prepare their organizations by educating themselves and their workforces about causal AI and building the interdisciplinary teams needed to develop applications.

Many companies today invest heavily in employee training on traditional ML and Generative AI models (such as ChatGpt) to remain competitive and innovative.If your organization intends to use Causal ML, it must include this technology in its AI literacy efforts.Employees who are attentive to the strengths and limitations of different approaches to AI will be able to find opportunities to use them effectively.

We have found that to excel at using Causal ML, teams need astrong experience in data science and machine learning,as well as sector knowledge. However, building these teams can be expensive, especially when companies need to hire data scientists or turn to external consultants and partners.

Additionally, data scientists and machine learning engineers are typically assigned to different teams. They must work closely when developing and implementing Causal ML models and have strong engagement with business stakeholders who have domain knowledge. (Domain knowledge is also essential in traditional machine learning, but is often applied less rigorously because teams don't fully consider the underlying relationships between variables when building models.)

For example, at the Neue Zürcher Zeitung, editors' and marketers' knowledge of editorial processes, customer preferences and long-term brand goals helps data scientists define the variables that measure these factors. At Hitachi ABB, engineers provide the information needed to define the manufacturing variables to include in the models.

Interdisciplinary teams are often plagued by a lack of common understanding, vocabulary and ways of working.Managers must foster an environment where cross-functional collaboration can thriveand in which all interested parties are involved in the model development process. Regular workshops, meetings, and training sessions where data scientists, machine learning engineers, and domain experts explore problems together, refine models, and discuss the implications of findings together can foster an environment where cross-functional collaboration thrives.

Machine learninghas changed the way many organizations make decisions;Causal ML can further deepen knowledge by predicting the effects of different choices on business results. Businesses are more likely to benefit from machine learning when decision makers trust the results. Knowing what Causal ML can do and how it compares to traditional ML can help you choose the right projects for each technology and increase your success rates.

When managers use Causal ML prudently to explore options for the simplest decisions, they can significantly improve their operations and, ultimately, their financial results.

Bibliography

S. Feuerriegel, Y.R. Shrestha, G. von Krogh, et al., “Bringing Artificial Intelligence to Business Management,” Nature Machine Intelligence 4, no. 7 (July 2022): 611-613; and P. Hünermund, J. Kaminski, and C. Schmitt, “Causal Machine Learning and Business Decision-Making“, SSRN, updated February 19, 2022,https://ssrn.com.

St. Feuerriegel, D. Frauen, V. Melnychuk, et al., “Causal Machine Learning for Predicting Treatment Outcomes,” Nature Medicine 30 (April 2024): 958-968; V. Chernozhukov, C. Hansen, N. Kallus, et al, “Applied Causal Inference Powered by ML and AI”, PDF file (published by authors on July 28, 2024), https:causalml-book.org; and C. Fernández-Loría and F. Provost, “Causal Decision-Making and Causal Effect Estimation Are Not the Same … and Why It Matters,” Informs Journal on Data Science 1, no. 1 (April-June 2022): 4-16.

M. von Zahn, K. Bauer, C. Mihale-Wilson, et al., “Smart Green Nudging: Reducing Product Returns Through Digital Footprints and Causal Machine Learning,” Marketing Science, Articles in Advance, published online August 8, 2024; E. Ascarza, “Retention Futility: Targeting High-Risk Customers Might Be Ineffective,” Journal of Marketing Research 55, no. 1 (February 2018): 80-98; J. Yang, D. Eckles, P. Dhillon, et al., “Targeting for Long-Term Outcomes,” Management Science 70, no. 6 (June 2024): 3841-3855; and M. Kraus, S. Feuerriegel, and M. Saar-Tsechansky, “Data-Driven Allocation of Preventive Care With Application to Diabetes Mellitus Type II,” Manufacturing & Service Operations Management 26, no. 1 (January-February 2024): 137-153.

G. von Krogh, S.M. Ben-Menahem and Y.R. Shrestha, “Artificial Intelligence in Strategizing: Prospects and Challenges,” in “Strategic Management: State of the Field and Its Future,” eds. I.M. Duhaime, M.A. Hitt and M.A. Lyles. (New York: Oxford University Press, 2021), 625-646.

“Premium chocolate production perfected: AI’s Role in Quality Excellence”, ETH AI Center, 11 December 2023,https://ai.ethz.ch.

J. Senoner, T. Netland, and S. Feuerriegel, “Using Explainable Artificial Intelligence to Improve Process Quality: Evidence From Semiconductor Manufacturing,” Management Science 68, no. 8 (August 2022): 5704-5723.

H. Wasserbacher and M. Spindler, “Machine Learning for Financial Forecasting, Planning and Analysis: Recent Developments and Pitfalls,” Digital Finance 4 (March 2022): 63-88.

J. Persson, S. Feuerriegel and C. Kadar, “Off-Policy Learning for Audience-Wide Content Promotions“, working document, 2023.

Senoner et al., “Using Explainable Artificial Intelligence,” 5704-5723.