What does an AI business case need to withstand CFO scrutiny?
The short answer
What matters is not the proposal itself. In several studies, the existence of a business case shows no demonstrable link to later success — but four other elements do: a named benefit owner, a quantified expected contribution, a stop limit set in advance, and a later comparison of realized against expected benefit.
The proposal is on the table, cleanly calculated, with a timeline and a promised benefit. It is approved. A year later the CFO asks what became of it, and the room starts arguing about attribution.
I have written such proposals myself and approved them myself. For a long time I thought the problem lay in the quality of the submission — too optimistic, too few scenarios. The research says something else, and it is uncomfortable.
The proposal itself says almost nothing about success
In a study of 200 investment initiatives, the practice “a business case was prepared” was removed from the explanatory model because it explained nothing statistically (Badewi 2016). In a second study of 71 initiatives, no link could be shown between the existence of a business case and later success (Holgeid and Jørgensen 2020).
Four other elements, by contrast, showed clear links there: a plan for realizing the benefit, a named assignment of responsibility for it, quantification of the expected benefit and its evaluation after completion. The strongest single link in the 2016 study was the named assignment.
A third study with 331 respondents in three countries supports the same point from the other side: of twelve practices examined, only one was associated with success in all countries — that actual results matched what the business case had anticipated (Serra and Kunc 2015). So what matters is not preparing the proposal. What matters is delivering on it later.
That shifts the work. The hours that go into the third decimal place of the return calculation are worse invested than the five minutes spent on the question of who owes the benefit.
Your estimate is narrower than it should be, and that has been measured
CFOs are not the cautious ones in this game. Itzhak Ben-David, John Graham and Campbell Harvey collected CFO forecasts over years and held them against actual outcomes. The executives each gave an interval in which the value should lie with 80 percent probability. They hit it in 36 percent of cases. In periods of high uncertainty, in 23 percent.
It becomes even clearer with their own projects. There, the same executives assumed a return volatility of just over five percent, while the historical volatility of comparable investments was above 36 percent. Confidence is greatest where one’s own project is being assessed.
For the submission, this means: a single expected value is not information, it is a stance. In my experience, a reviewer expects three scenarios anyway.
In IT initiatives the problem is not the average but the tail
Bent Flyvbjerg and colleagues examined 11,011 projects of 23 types for cost overruns. Only 41 percent of IT projects exceed their budget at all. Those that do, however, reach several hundred percent at the upper tail of the distribution — on average over 450 percent within that group.
Mathematically, the consequence is unpleasant: for this type of project there is no reliable average, because the distribution has no stable mean. The question “what does something like this usually cost” cannot be answered. Only this can be answered: how much are we willing to lose, and how will we notice that we have got there.
For AI initiatives, that is exactly the relevant question, not the point estimate. The market side shows that value contributions remain thin: 37 percent of companies attribute any effect on operating profit to AI at all, and this share has not grown compared with the previous year (McKinsey 2026, 1,719 participants). In the third quarter of 2026, KPMG found among 2,131 executives that 89 percent report measurable value — and 12 percent regularly set this value against the costs.
A limit set in advance works, a committee does not
The largest synthesis on why failing initiatives keep getting funded covers 166 studies (Sleesman and colleagues 2012). Two of its findings are counterintuitive. The strongest driver is not the amount of costs already sunk. It is proximity to completion. And decision authority spread across several people intensifies escalation instead of slowing it. A committee offers no protection.
What works is narrower. In experiments with managers, a decision rule specified in advance slowed escalation; a self-set rule barely did, and additional information barely did (Boulding, Morgan and Staelin 1997). In a further study, a minimum target set in advance reduced continued funding, while the instruction to decide more carefully had no effect (Simonson and Staw 1992).
One caveat belongs here: where limits are drawn sharply, the incentive to delay bad news rises. The limit therefore needs a date on which it is measured — not a follow-up question someone has to ask.
What belongs in the submission so it survives scrutiny
Four elements, and none of them needs a model.
First, a figure that already exists in the company, and the share of it this initiative is meant to contribute. Second, a name: the person who owes the benefit. That is rarely the project lead who runs the implementation. Third, the limit: at which value by which date the initiative ends. Fourth, a date for comparing the realized with the expected contribution.
Plus one adjustment to review practice. Among German CFOs, about half use the payback period, just under half net present value, somewhat fewer the internal rate of return (Brounen, de Jong and Koedijk, 132 respondents). The survey is more than twenty years old and the only robust one for Germany. Presenting only a net present value misses half of the reviewers.
This is the part BizzPlAI is built for as a Strategy Execution System: the expected value contribution, the goal and the stop condition remain retrievable from the decision onward, until the date of the comparison. The estimate itself does not get better as a result. It only becomes verifiable.
A test for the next proposal on your desk: strike out the return calculation and read the rest. If no name, no limit and no date remain, you have not read a business case. You have read a declaration of intent with appendices.
Common follow-up questions
Why is a business case alone not enough?
In a study of 200 initiatives, the practice "business case prepared" was removed from the explanatory model for lack of statistical significance. What remained was the named assignment of responsibility for the benefit. The document does not decide; the assignment does.
Which appraisal methods does a CFO in a mid-sized company expect?
According to the most recent robust survey of German CFOs, about half use the payback period, just under half net present value and somewhat fewer the internal rate of return. Presenting only a net present value reaches only part of the reviewers.
Why is the stop limit more important than the estimate for AI initiatives?
Because cost overruns in IT projects are extremely unevenly distributed. Only a minority overrun at all, but at the upper tail the overruns reach several hundred percent. No better point estimate protects against this distribution — only a limit does.
Sources
- Global AI Pulse Q3 2026 — survey of 2,131 executives, fieldwork July to August 2026KPMG International, 2026
- Realizing benefits in IT projects: practices and their relation to success, IEEE Conference on Business Informatics — 71 initiativesKnut Kjetil Holgeid and Magne Jørgensen, 2020
- The impact of project management and benefits management practices on project success, International Journal of Project Management 34(4) — 200 initiativesAmgad Badewi, 2016
- Benefits Realisation Management and its influence on project success and on the execution of business strategies, International Journal of Project Management 33(1), pp. 53–66 — 331 respondents in three countriesCarlos Eduardo Martins Serra and Martin Kunc, 2015
- Managerial Miscalibration, Quarterly Journal of Economics 128(4) — CFO forecasts against actual outcomesItzhak Ben-David, John R. Graham and Campbell R. Harvey, 2013
- The Empirical Reality of IT Project Cost Overruns, Project Management Journal — 11,011 projects in 23 typesBent Flyvbjerg et al., 2025
- Cleaning Up the Big Muddy: A Meta-Analytic Review of the Determinants of Escalation of Commitment, Academy of Management Journal 55(3) — 166 studiesDustin J. Sleesman, Donald E. Conlon, Gregory McNamara and Jonathan E. Miles, 2012
- Pulling the Plug to Stop the New Product Drain, Journal of Marketing Research 34(1) — experiments with managersWilliam Boulding, Ruskin Morgan and Richard Staelin, 1997
- The Effect of Prior Experience and Current Decision Context on Escalation, Organizational Behavior and Human Decision ProcessesItamar Simonson and Barry M. Staw, 1992
- The state of AI — survey of 1,719 participants in 97 countriesMcKinsey & Company, 2026
- Corporate Finance in Europe: Confronting Theory with Practice — 132 German CFOsDirk Brounen, Abe de Jong and Kees Koedijk, 2004
See how strategy arrives.
In 30 minutes, we show you how to steer your initiatives instead of merely managing them.