Every AI business case someone hands me starts the same way. A number gets picked before anyone’s agreed on what the number is supposed to represent. Revenue? Cost avoided? A risk the company no longer has to carry? Nobody’s agreed, yet the number got picked anyway.

Like most product or technology ventures, what “good” is needs to be defined by someone and rarely can be inherited.

The math is genuinely different every time. Whether it’s different cost structures, different appetite for a multi-year payback, or something else, what’s constant isn’t the formula. It’s the discipline underneath it, and that discipline has three steps I keep seeing skipped everywhere, from business cases to Reddit.

Picking the Bucket Before the Metric

We reach for a “metric” before anyone’s agreed what “the return” is supposed to mean. Value from AI work almost always lands in one of four places.

1. Revenue created
2. Customer kept (churn reduction)
3. Reduced operational costs
4. Reduced risk

Pick the bucket first and the metric follows from that choice. Skip straight to a number, any number, and a genuinely good initiative gets measured against the wrong yardstick, then defended badly in the room where it matters. I’ve watched a great result get presented in the wrong bucket (a retention win dressed up as an efficiency stat) and watched it land flat because the room was evaluating it against a question it never answered.

A Gartner survey of finance leaders, published in July 2026, shows how common that mismatch actually is. Just 20% of finance AI projects lean toward decision quality, the outcome boards say they value most. 45% lean toward straightforward productivity gains instead. Boards want growth and sharper decisions. Finance keeps investing in productivity instead. That’s the same gap I keep running into when a business case reaches the CFO’s desk. A genuinely good initiative, measured against the wrong bucket from the start.

The Learning Curve Is Part of the Budget

Once the bucket’s picked, most business cases still assume value starts on day one, as though flipping the tool on were the finish line instead of the starting gun. It isn’t. Before a developer builds anything, someone has to source and clean the data, then do the unglamorous data science work of testing and tuning the model against edge cases it hasn’t seen yet. That’s what “building” actually means with AI, and it’s a real, budgetable cost, not a rounding error on the schedule.

Deloitte’s 2025 survey of more than 1,800 executives across Europe and the Middle East found typical AI payback landing at two to four years, several times longer than the seven to twelve months teams typically expect from a conventional tech rollout (a benchmark Deloitte itself cites from separate industry research, not from this survey). Only 6% of the executives Deloitte surveyed saw payback inside a year. Deloitte’s own researchers call the finding “directionally applicable globally,” even though the underlying survey never left Europe and the Middle East.

A January 2026 BCG survey of executives across 16 markets suggests the caution travels further than that. Confidence that AI investment will actually pay off lands at 52% among US CEOs, 44% in the UK, and 61% in Europe, against roughly three-quarters of CEOs in India and Greater China. The skepticism isn’t a European quirk. It’s an uneven band any US-based business case still has to plan around.

All of that data work is sunk cost the moment it’s spent, which is exactly why the bucket from the first section matters before, not after, the budget clears. Picture a team standing up a support model specifically to reduce churn. Most of the months that went into it were data collection, cleaning, and tuning that never shows up as its own line item. If the model gets an answer wrong even occasionally, the customer doesn’t quietly churn. They call. A churn-reduction bet can turn into an operational-cost problem the business case never budgeted for, spending against a bucket nobody picked. Setting the payback horizon honestly up front, instead of backloading the case onto whichever fiscal quarter needs a win, is what keeps that sunk cost from becoming a surprise. I got this wrong once in a fintech role, treating the learning curve as a cost to minimize rather than a sunk cost to plan for.

Surviving a Budget Review

Bucket and horizon still don’t matter if the case gets presented in the wrong dialect. Getting budget approved for a tax-calculation platform meant walking into a room with people who track cost per return processed and audit exposure for a living. They didn’t care how the matching logic worked. They cared whether it moved margin, lowered a cost, grew revenue, or protected a retention number they already watched every quarter.

That’s the whole translation. Any CFO is already tracking revenue, cost, retention, and risk before you walk in, the same four buckets from a page ago, just spoken in the vocabulary the room already uses. Two more concepts round out that vocabulary. Amplification is getting more output without adding proportional cost, the same team handling more volume. Repeatability is whether this is a one-time win or a capability the business can lean on again next quarter without rebuilding it from scratch.

Picture a contact-center team translating the exact same win two different ways. A 30-second reduction in average handle time means very little to a CFO sitting outside that team. Reframed as handling meaningfully more call volume without adding headcount, the hire that reduction quietly avoided, it’s the same operational gain, priced in the only language that survives a budget review.

I’m still working out how much of this translation can be taught versus how much just gets lived through a few bad meetings first. What I’m fairly sure of is that skipping any one of these three steps doesn’t make the underlying case wrong. It just leaves it unfinished, and an unfinished case reads as a wrong one in the room where it counts, the same gap I keep running into when translating product reality back up into investment decisions.

Which of those four buckets (revenue, retention, cost, or risk) does your current AI initiative actually belong to, and has anyone in the room actually agreed on that, or has everyone just assumed?