Ask any CIO how long it took to get their most recent generative AI budget approved, and the answer is almost always the same: not long. Boards have stopped debating whether to fund a chatbot or a copilot. What they have started debating, two quarters later, is what that spending actually delivered, and increasingly, the answer is uncomfortable.
The honest answer is usually, not much.
IBM’s 2025 CEO study, drawn from 2,000 executives, found only 25% of AI initiatives delivered the ROI leadership expected, with 16% scaling beyond a pilot. MIT’s Project NANDA went further: 95% of enterprises deploying generative AI saw zero measurable impact on profit or loss. This is not AI underdelivering. It is enterprises funding the wrong layer of the stack. Most AI budgets go toward visible functions like sales and marketing content, while the highest measurable returns sit in back-office decision automation that almost nobody funds. Businesses are paying to describe their problems faster. Very few are paying to understand what is causing them, which is the layer FireAI has chosen to build.
None of this means enterprises are pulling back. Gartner’s July 2026 forecast puts worldwide AI platform spending at $64.2 billion for the year, a 63% jump from 2025. Boards have not lost faith in AI. They have lost patience with generative AI as the entire strategy.
Why Generative AI Cannot Do What Causal AI Does
The difference sits at the level of what each system is built to do. A large language model is, at its core, a correlation engine trained on which words tend to follow which other words. Ask it a business question, and it produces the most statistically plausible answer based on patterns it has seen before. That is useful for writing and searching. It is not the same as understanding why something happened.
TheCUBE Research put this precisely in a January 2026 report: “Data is not knowledge: statistical correlation and dynamic retrieval do not create understanding of what causes what and why. A prediction is not a judgment: forecasts do not decide; decisions require trade-offs, constraints, and consequences.” That distinction is the practical reality behind the entire argument for causal AI.”
Judea Pearl’s causal hierarchy gives this some precision. At the base level, association, a system can tell that two variables move together without knowing which one is driving the other. Ice cream sales and drowning rates both rise in summer, but neither causes the other; heat does. Generative AI operates almost entirely at this level. Real decision-making requires climbing higher, to intervention, predicting what happens if you deliberately change one variable, and counterfactual reasoning, working out what would have happened had a past decision gone differently. Causal AI is built for both, mapping cause and effect as a network of relationships rather than a black box of statistical weights.
The Enterprise AI Stack Has Three Layers, Not One
CIOs getting this right have stopped treating enterprise AI as one technology. The interface layer is generative AI: chat, copilots, summarisation. The reasoning layer is causal AI: the system that works out why a number moved and what happens if a specific action is taken. The action layer is agentic AI, executing decisions autonomously against real business systems.
The sharpest risk sits where businesses skip the reasoning layer entirely. In April 2026, a coding agent working for a car-rental software company was given a simple task: fix a piece of code. Nothing in its training told it that deleting the company’s production database, and every backup along with it, was an unacceptable way to get the job done quickly. So it did exactly that, in seconds. Nobody hacked the system. Nobody told the agent to cause damage. It simply had speed as its only goal and no understanding of what it was destroying to get there. A May 2026 review of over 7,200 public AI incidents found 188 separate cases of this kind of agent-inflicted damage, none involving an external attacker. Agentic AI without causal reasoning underneath it is a fast, tireless system with no understanding of what it is allowed to break.
The Governance Case Is No Longer Optional
This is where it becomes personal. IBM’s June 2026 study of 2,000 technology executives found 66% of CIOs and CTOs are now held personally accountable for AI systems they do not fully control. That figure should change how every technology leader evaluates a build-versus-buy decision on AI.
Generative AI systems are structurally hard to audit. Billions of parameters sit between an input and an output, and nobody can fully explain why the model produced one answer over another. Causal systems map exactly which variables were considered, which were ruled out as confounders, and which assumption drove the recommendation. When something fails, an auditor traces it to a specific node in the reasoning chain rather than retraining an entire model and hoping the problem does not recur.
Gartner published its first Magic Quadrant for Decision Intelligence Platforms in January 2026, defining the category as software that combines data, analytics, and AI to model, orchestrate, and govern decisions end to end. A category that did not formally exist two years ago is now something Gartner rates vendors against.
The Adoption Gap Is Real, and It Is an Opportunity
Gartner’s Hype Cycle places overall decision intelligence adoption at 5% to 20% of the addressable market, with mainstream maturity two to five years out. Enterprises that have pushed causal AI into genuine production sit closer to 5%, based on how few generative AI deployments reach measurable business impact.
That gap is not evenly spread. Banking leads, largely because credit scoring and anti-money laundering detection already carry explainability requirements black-box models cannot satisfy. Healthcare follows, with causal models now matching the diagnostic accuracy of the top quartile of physicians on complex cases. Manufacturing reports downtime cuts of 15 to 30% where causal models replace standard predictive maintenance, because they find the actual mechanical root cause rather than a correlated symptom.
India sits in an interesting position. Enterprise AI hiring here has already shifted away from generic generative AI experience toward model validation and decision intelligence specifically. The gap most Indian enterprises face is not appetite. It is data readiness, and that is where transformation budgets should be going now.
What This Means for the CIO’s Next Budget Cycle
The practical takeaway is not complicated, even if the mathematics underneath it is. Audit the AI stack in production against three questions: can it reason about why an outcome occurred, can it explain that reasoning to someone defending the decision to a regulator or a board, and does it act only within boundaries that reasoning has validated. Any agentic system failing the third question should not have autonomy yet, regardless of how impressive its interface feels.
Causal capability should not be a feature bolted onto a generative AI programme. It belongs underneath every AI initiative the enterprise runs, the way a database sits underneath every application. The organisations that understood this before their competitors are already showing up in Gartner’s Leader quadrant. Everyone else is still funding the interface layer and wondering why the return has not materialised.
Authored by Vipul Prakash, Founder & CEO of FireAI