The Business Case for AI Is Almost Always Wrong

I've reviewed a lot of AI business cases in the last two years. They all look roughly the same. A clean spreadsheet. A plausible narrative. Some version of: we'll automate X process, save Y hours per week, reduce Z headcount or error rate, and achieve ROI within 18 months.

The math always works on the slide. The math almost never works in production.

Not because the teams building these cases are sloppy or dishonest. They're usually sharp people doing their best with available information. The problem is structural. The standard business case format — the one every org uses, the one every consulting firm teaches — was designed for deterministic investments. Buy a machine, it produces widgets at a known rate. Hire a team, they ship a product on a scoped timeline. The inputs are knowable. The outputs are predictable.

AI doesn't work like that. And the business case format doesn't have a line item for "things we can't predict yet."

The four assumptions that always break

Every AI business case I've seen depends on at least four assumptions that are treated as given but are actually open questions.

The data will be ready. The business case assumes the model will be trained or fine-tuned on clean, representative, well-labeled data. In practice, the data is scattered across three systems, maintained by teams with different schemas, and nobody has validated it end-to-end. The data preparation work — the unglamorous pipeline construction, the schema reconciliation, the quality auditing — takes three to six months longer than anyone estimated because nobody estimated it in the first place. It wasn't in the business case. It was assumed.

The team will stay dedicated. The business case assumes a cross-functional team focused on the AI initiative: a PM, a data scientist, an engineer, someone from the business unit. In reality, these people have day jobs. The data scientist gets pulled onto an executive fire drill. The engineer gets reassigned when another project hits a deadline. The business unit representative attends two meetings, then delegates to someone junior who doesn't have decision-making authority. By month three, the "dedicated team" is two people working on it part-time.

Integration will be straightforward. The business case has a line for integration but it's a fraction of the total budget. It assumes the AI system will connect to existing workflows cleanly. It won't. The ERP system has an API that hasn't been updated in three years. The CRM has custom fields that nobody documented. The downstream process that's supposed to consume the AI output was designed for human decision-making and has approval gates that don't map to automated outputs. Integration isn't a line item. It's a project unto itself.

Users will adopt it. The business case assumes that once the system is deployed, people will use it. They won't — at least not the way the business case imagines. Some users will distrust it and verify every output manually, adding a step instead of removing one. Some will find it easier to keep doing things the old way. Some will use it selectively, cherry-picking outputs that confirm their existing judgment. Adoption isn't deployment. It's behavior change, and behavior change takes sustained effort that isn't in the budget.

The costs that don't make the slide

The business case captures the visible costs: licensing, compute, development hours. It misses the costs that are harder to quantify but often larger.

Change management. Getting an organization to actually change how it works is expensive. Not in licensing fees — in leadership attention, training, workflow redesign, and the productivity dip that happens during transition. Most business cases allocate zero dollars and zero hours for this. The implicit assumption is that people will adapt because the new tool is better. That assumption has a poor track record.

Monitoring and maintenance. The model works today. Will it work in six months when the input data distribution has shifted? When a source system gets updated? When a new product line introduces categories the model hasn't seen? Production AI systems need ongoing monitoring, periodic retraining, and someone who's paying attention. That's operational cost. It doesn't end when the project "launches."

Organizational friction. The meeting where the legal team asks about liability. The week where the compliance team reviews the model for regulatory risk. The month where the data governance team discovers the training data includes PII that wasn't supposed to be there. These aren't edge cases. They're the normal course of deploying AI in a large organization. They consume time, require decisions, and slow everything down. The business case treats the timeline as a straight line. The actual timeline has delays nobody planned for.

What an honest business case looks like

The organizations I've seen succeed with AI aren't the ones with the most ambitious business cases. They're the ones with the most honest ones.

An honest business case includes a section on what has to be true for the numbers to work — and an assessment of whether those things are actually true today. It includes a realistic integration estimate that accounts for the state of existing systems, not the state described in vendor documentation. It includes change management as a funded workstream, not an afterthought. It includes a post-deployment operating cost that covers monitoring, maintenance, and iteration.

Most importantly, an honest business case includes a kill criteria. Under what conditions would we stop this project? What would the data have to show? If you can't answer that question before you start, you're not making a business case. You're making a commitment, and commitments are hard to reverse even when the evidence says you should.

The business case that gets funded is the one that tells leadership what they want to hear. The business case that delivers value is the one that tells leadership what they need to hear. Those are rarely the same document.

The slide deck is not the plan. The plan is what you do when the slide deck meets reality.