Why Most AI Business Cases Fail the CFO Test
The typical AI business case presented to a CFO or board looks like this: a slide with a large potential productivity gain, a software license cost, and an implied ROI. What it omits is everything the finance team will ask about: change management, data readiness, integration work, ongoing operations, and the timeline to get from pilot to a state where the productivity gain is actually realized.
Research from the BCG and MIT Sloan Management School published in 2023 provides one of the better-controlled studies of AI impact on knowledge workers. The Dell'Acqua et al. study found that consultants using AI improved their performance on a defined task set by 12.2% on average. That is a real and meaningful number, but it comes from a controlled experiment on a narrow task type. The gap between controlled study results and enterprise-scale deployment reality is where most business cases fall apart.
The McKinsey Global Institute's June 2023 analysis of generative AI's economic potential estimated that the technology could deliver productivity equivalent to trillions of dollars annually across industries. That framing is useful for setting strategic context. It is not a business case for your specific pilot. Your business case needs to be grounded in the specific process you are changing, the specific cost structure of that process today, and the specific changes in labor time or output quality the AI will produce.
The ROI Framework for AI Initiatives
AI ROI follows the same basic logic as any capital investment, with one important adjustment for the cost side: total cost of ownership for AI includes categories that traditional IT projects often exclude.
The framework is: ROI = (Value Created - Total Cost of Ownership) / Time to Value. Each term requires specific definition.
Value Created is the dollar-denominated benefit from the AI deployment. It comes from three sources: labor time saved (hours reduced multiplied by fully-loaded cost per hour), quality improvement (error reduction, faster cycle time, better output, measured in downstream cost or revenue impact), and new capability (tasks the organization could not do before, valued at what you would have paid for the equivalent output).
Total Cost of Ownership is where most business cases undercount. The five components are: software and compute (license fees, API costs, cloud infrastructure), data engineering (the work to prepare, clean, and maintain the data the AI system needs), integration (connecting the AI to the systems where the work actually happens), change management (training, workflow redesign, and the productivity dip during transition), and ongoing operations (monitoring, evaluation, retraining, and the people who keep the system working after launch).
Time to Value is the elapsed time from project start to the state where the value is actually being realized at scale, not the state where the pilot is working on a small sample.
The Five Components of Total Cost of Ownership
The cost category that executives most consistently underestimate is data preparation and cleaning. Before any AI system can work reliably on your business problem, the data it needs must be in a form the system can use. For most enterprise use cases, this means structured retrieval, consistent formatting, access controls, and freshness guarantees. In organizations where data is spread across legacy systems, inconsistently maintained, or held in formats that require manual extraction, data engineering can represent more than half the total project cost and most of the timeline risk.
The fifth component, ongoing operations, is the cost that surprises executives after deployment: the people and processes needed to monitor performance, catch drift, evaluate edge cases, and keep the system aligned with evolving business rules. AI systems are not "deploy and forget" infrastructure. They require ongoing human oversight, particularly as the data distribution shifts over time.
Three Failure Modes That Kill AI Investments
Over-scoping the MVP. The most common failure mode is designing a first deployment that tries to solve the entire problem rather than one well-defined slice of it. An AI pilot that is scoped to handle all customer inquiries across all product lines and all languages will take longer, cost more, and produce ambiguous results that make it hard to learn anything. A pilot scoped to handle one specific inquiry type for one product line produces clear signal about whether the approach works and scales. Start narrow, learn fast, expand deliberately.
Ignoring the change management budget. Technology adoption requires behavioral change. The people who currently own the process the AI is changing must understand why the change is happening, what it means for their roles, and how to use the new system effectively. Organizations that treat AI deployment as a technology project rather than a change project consistently see lower adoption, more workarounds, and weaker realized value than the business case projected.
Underestimating data readiness costs. This failure mode appears in two forms. The first is the "our data is in good shape" assumption that collapses during the first sprint when the team actually touches the data. The second is the assumption that data readiness is a one-time cost that ends at launch. In practice, data quality requires ongoing maintenance as business processes change, new data sources are added, and the AI system's coverage expands.