How to build an AI business case before the ROI exists

Most AI business cases answer the wrong question. They argue for buying AI, when the money to buy AI has usually already been spent, and spent again: by Deloitte's 2025 read, the large majority of organizations increased AI investment last year and plan to increase it again, while the returns stay stubbornly thin. The harder, more honest question, the one finance is actually asking in 2026, is why the AI you already bought isn't paying off. A widely cited MIT study found that 95% of enterprise generative-AI efforts have produced no measurable impact, and named the cause a learning gap: the tools never got integrated into how people actually work. So the gap your AI business case has to close isn't a tooling gap. It's the distance between the licenses you own and the work that hasn't gotten any better.

That reframes the whole document. You're not asking the company to bet on AI. You're asking it to fund the one thing standing between approved spend and real return: capability.

Buying a seat is not the same as someone using it. By IBM's 2026 CEO research, around 85% of employees have access to AI tools at work while only about a quarter use them regularly, a gap executives now rank above cost and security as their top AI concern. Every one of those idle seats is money already spent earning nothing, which means the cheapest, fastest lever on AI ROI is not another platform; it's getting the people you've already licensed to actually use what they have on real work. That's the lever an AI transformation is supposed to pull, and it's the one most rollouts skip. Tool spend buys potential. Capability spend is what converts it, and only one of the two usually got funded.

Here's the objection a good CFO will raise, and it's fair: soft productivity gains don't survive a finance review. "We trained people and they feel faster" is exactly the kind of claim that gets discounted, and rightly, because self-reported gains often shrink under scrutiny. Finance approves hard ROI, labor hours saved, cycle time cut, error rates down, and it wants a payback window measured in this fiscal year, not in three. AI investments that promise returns three to five years out make finance nervous for good reason. The answer is not to wave the objection away with vision. It's to refuse to bring a soft number in the first place. Measuring the return after the fact is its own discipline, and we cover it in how to measure AI ROI; this piece is about justifying the spend before that return exists, by making the capability claim hard rather than soft.

Capability spend stops being soft the moment you tie it to a specific, measured workflow. Pick one piece of real work, proposals, reconciliations, support tickets, and baseline it before anyone is trained: how many hours, how many errors, how long a cycle. Then train the people who own that work, on that work, and measure the same numbers again. The evidence that this moves the needle is strong: an LSE study with Protiviti found trained employees save about 11 hours a week against 5 for the untrained, and 93% of trained employees use AI regularly versus 57% of those left to figure it out, while 68% had received no AI training in the past year. Training, not tool access, is what turns a license into usage, and usage into a number you can defend. Anchor it in the roles and team that own the workflow, and the claim writes itself.

Skip the generic four-box template every consultancy hands out. The case that gets funded is narrower and harder: the one workflow you'll prove it on, the baseline you measured, the capability spend required, the result you expect inside the fiscal year, and the cost of doing nothing, which is the idle-license bill plus the work that stays slow. Naming the do-nothing cost is what turns this from a request into a comparison, and it's usually the most persuasive line on the page. Done this way, the business case isn't a bet on AI; it's a plan to collect on an investment the company already made. That's also how AI adoption earns its next round of budget: not on a promise, but on a proven number you can repeat in the next function.