How Causal Reasoning Could Improve Enterprise AI Adoption

I have been writing for several years now about the benefits of using artificial intelligence in the workplace. And for several years, I have been highlighting one very important statistic: only 20% of companies are able to scale their use of AI. The vast majority of companies are stuck in what I call the “prediction trap”, they are able to create very accurate predictions with their AI models, but those models do not support any decisions at all. Causal reasoning for the win! The traditional way of doing predictive analytics is to come up with the best model for historical data. However, this approach is completely blind to the underlying mechanisms that produced the data in the first place. Causal what-if analysis on the other hand, is able to tell us why something happened in the past and what would happen in the future under different circumstances. Going back to my previous example, a predictive model might find a correlation between a company’s marketing spend and their revenues. However, using causal analysis, one can see that the seasonality of a company’s customers is the cause of both the marketing spend and the revenues. Therefore, spending more money on marketing in order to gain more revenues would be almost worthless. Counterfactuals: Reverse the Situation and Get Answers to Reverse Questions. Often people refer to prediction models that forecast future events and situations. These situations are then anticipated and in the end they do not bring any added value. In order to be able to make interventions in situations (and to be able to predict the effects of these interventions) it is better to reverse track to a point in time in the past. In that situation, and from that point in time it is possible to ask reverse questions (what if questions) and forecast what would have happened if the situation would have developed in a completely different way. This type of analysis is also known as causal analysis. A large number of frameworks exist to analyze and answer these types of questions in a causal manner. Examples of these frameworks are the Rubin Causal Model (RCM), Pearl’s Structural Causal Models with Directed Acyclic Graphs (SCMG) and the Abduct-Act-Predict (AAP) procedures. In recent years, also ways have been developed to incorporate causal structures into deep learning models in order to gain insights in a data driven fashion, while at the same time generating accurate predictions. The key areas of value in this emerging technology are numerous, including: CUBE Research predicts that causal AI decision intelligence will be an emerging priority for the enterprise in 2026; Followed by 69% of the enterprise AI employees that CUBE Research surveyed will adopt causal AI in 2026. A couple of weeks ago, I wrote that the analyst role is evolving from being passive recipients of predictions generated by models, prescribing of interventions. In order to be able to do this effectively, the analyst must first be able to communicate to stakeholders the assumptions made by the model that generated predictions. This can be achieved through the analyst communicating causal diagrams, as well as through the creation of ‘what-if’ narratives of possible interventions that can be taken. The analyst also needs to be able to challenge and verify the work of the model. Adoption of Causal Reasoning However, there are many obstacles for adopting causal reasoning for AI-driven decisions. There is a huge “faithfulness gap” for the current generation of LLMs, i.e. only 26% on average for the studied explanations was classed as “very faithful” by humans, followed by 48% that were classed as “roughly faithful” (with an average score of 74% across all studied LLM explanations, CMU study). This means that in the vast majority of cases the model’s attempt to provide explanations for its outputs will not actually correspond to the real causes that led to those outputs and hence this will create a huge governance problem for using AI systems in critical decision making contexts. In addition, many other large barriers to adoption, such as security (52%), integration (48%) and many large skill gaps for both statistical modeling as well as for causal modeling exist. Causal AI Market size will grow from $56.2 million in 2024 to $456.8 million by 2030, at a Compound Annual Growth Rate (CAGR) of 41.8% in the forecast period. | Year | Market Size ($ Million) | YoY Growth | |----|----|----| | 2024 | 56.2 | — | | 2025 | 69.6 | 23.8% | | 2026 | 102.5 | 47.3% | | 2027 | 153.9 | 50.1% | | 2028 | 223.0 | 44.9% | | 2029 | 314.0 | 40.8% | | 2030 | 456.8 | 45.5% | \ Conclusion The bottom line for me is that causal reasoning is the missing layer of intelligence in AI automation to become Decision Intelligence that enterprises can trust. It already matters and as 62% of enterprises plan to move to Decision Intelligence in the next 18 months or so, it will become increasingly important for analysts to master the art of using counterfactuals (i.e. asking and answering questions such as “what would happen if…?”). That new layer of intelligence requires a new organization around it. And that organization is already underway. References https://www.marketsandmarkets.com/Market-Reports/causal-ai-market-162494083.html https://thecuberesearch.com/why-causal-ai-decision-intelligence-2026/ https://www.nature.com/articles/s43588-025-00814-9 https://www.microsoft.com/en-us/research/publication/a-causal-ai-suite-for-decision-making/ https://hbr.org/sponsored/2026/02/close-your-workforces-ai-skills-gap-by-designing-an-adaptive-organization https://xchange.avixa.org/posts/the-real-barriers-to-ai-adoption-aren-t-what-you-think \