The budget round at a regional hospital group is nearly over. Twelve AI proposals are on the table, and one stands out by a distance: an optimiser that rebuilds operating room schedules when the elective backlog gets out of hand. Its business case is three times larger than the next one. The board is ready to approve it first, because if any AI business case deserves the money, surely it is the biggest one.
That instinct is understandable, and it is often wrong. In the MetroMinds Library the use cases with the highest value are, on average, the hardest to deliver and the riskiest to run. The size of a business case tells you how much could be gained. It tells you very little about how likely you are to get it, which is why judging AI ROI on that one figure is risky.
Big value and easy delivery rarely come together
Every use case in the MetroMinds Library carries a business value score and a feasibility score, both on a scale of 1 to 5. Of the 4,980 use cases, 803 get the top value score of 5. Only 26 of those 803 also get the top feasibility score. That is about 3 percent.
The averages tell the same story. Across the whole Library the mean feasibility score is 3.3. For the top-value group it drops to 2.87, the lowest of any value level. More than a third of the top-value use cases (34 percent) score 2 or lower on feasibility. In the next group down, with a value score of 4, that share is 25 percent.
Risk rises with value as well. The mean risk score of the top-value group is 4.30, against 3.82 for the Library as a whole. That makes sense once you look at what these use cases do. They tend to touch scarce resources, clinical or financial decisions, or several systems at once. The same reach that creates the value also creates the integration work, the governance questions and the cost of a mistake.
Why the largest number attracts the least scrutiny
Large business cases have a natural advantage in a steering committee. They are usually written by the most ambitious sponsor, they promise a story the board likes to hear, and they make smaller proposals look timid. Questions tend to focus on the timing of the benefit rather than on whether the benefit will arrive at all.
This is a well-known pattern, and it is not limited to AI. HM Treasury's guidance on business cases states it bluntly: "Project appraisers have the tendency to be over optimistic." [1] It asks for explicit adjustments to costs, benefits and duration, based on data from past or similar projects. Lovallo and Kahneman described the same effect in the Harvard Business Review: executives build scenarios of success and overlook the potential for mistakes and miscalculations [2].
AI adds its own twist. Because many use cases are new to the organisation, there is often no internal track record to correct the optimism. The business case leans on vendor material or on the best result from a pilot. That is exactly where a from comparable implementations helps, as long as you use the whole range and not only the top of it.
The range matters more than the headline
Every benchmark in the Library comes as a range with a low, a mid and a high value. For the typical use case, the high end of the range is twice the low end. That holds for the top-value group too: in 51 percent of those use cases, the high benchmark is at least double the low one.
In practice this means the same use case can support a modest business case or a spectacular one, depending on which end of the range the author picks. A sponsor who uses the high end and leaves out scope and adoption is not lying. They are simply presenting the best outcome as the expected one. The box below shows how far apart the two versions can be.
What feasibility hides
A low feasibility score rarely means the idea is bad. It usually means the organisation is not ready to deliver it yet. In the Library, feasibility goes down when integration is hard, when data access is limited or untrusted, and when the orchestration across systems is complex. Those are the conditions that turn a twelve-week project into an eighteen-month programme.
That matters for the business case, because delay is a cost the original number rarely includes. A use case that promises 500,000 euros a year but needs two years of data work before it delivers anything may be worth less, in present value, than a 150,000 euro use case that runs next quarter. It also ties up the scarce people who could have delivered three smaller wins in the meantime.
Low feasibility also lowers confidence in the number itself. Benefits that depend on clean data and smooth integration are the first to shrink when either turns out worse than planned.
How to read a big AI business case
None of this means you should avoid the big ideas. The top-value use cases are often the ones that change how a service runs. The point is to fund them with your eyes open. A few questions help.
- Which end of the benchmark range does this business case use, and why that end?
- Are scope, adoption and confidence factors applied, and who agreed them?
- What has to be true about data, integration and governance before the first euro arrives?
- How long until that is true, and what does the delay cost?
- What would a smaller first step look like, one that proves the assumptions before the full investment?
In the MetroMinds value and feasibility matrix, many of these proposals land in the quadrant called : strong value, but harder to realise. That is a perfectly good place for a use case to be, as long as everyone treats it as a bet. It means staged funding, clear evidence before each step, and the willingness to put a use case on hold when the evidence does not arrive. Proposals in the Scale now quadrant, with strong value and strong feasibility, are rarer and deserve to move first.
If you want to see why several confident business cases can still add up to the wrong total, our article on AI business cases that count the same savings twice looks at the portfolio side of the same problem.
How MetroMinds helps you weigh value against delivery
In MetroMinds, every use case in your portfolio shows its business value next to its feasibility, risk and confidence, so the biggest number never travels alone. The business case uses the low, mid and high benchmark ranges from the Library and makes scope, adoption and confidence explicit, so the board can see which assumptions carry the figure.
The value and feasibility matrix then places each use case in one of four quadrants, from Scale now to Park / rethink. That gives you a shared language to say yes to a strategic bet without pretending it is a quick win.
Want to see how your largest AI proposals score on feasibility? Book a demo
Sources
- HM Treasury, "Green Book supplementary guidance: optimism bias", 21 April 2013.
- Dan Lovallo and Daniel Kahneman, "Delusions of Success: How Optimism Undermines Executives' Decisions", Harvard Business Review, July 2003.
- MetroMinds Library, 4,980 AI use cases with benchmark ranges based on real-world implementations, August 2026.
