We have sat on both sides of the table. We have presented cases for data and AI investment, and we have watched other people present them. Most of them fail, and the failure usually takes less than ten minutes.
It is rarely the idea that dies. It is the case.
Here is the pattern. Someone presents a slide with a large number on it. The number is a benefit: cost saved, revenue created, hours returned. A board member asks where the number came from. The answer is a hedge: an industry benchmark, a vendor's estimate, an internal workshop where everyone agreed it felt about right. A second board member asks what happens if the number is wrong. There is no answer to that, because the case was never built to be wrong. It was built to be approved.
That is the tell. A credible case is built to be tested. A vague one is built to be believed.
After enough of these meetings, you start to see that the credible cases share a handful of properties, and none of them are about the size of the number.
The baseline is stated, and it is frozen
Not "we currently spend roughly", but the actual figure, measured before anyone asks for money, written down where nobody can quietly improve it later. If the baseline moves after approval, the benefit becomes unmeasurable and everybody knows it. A case without a fixed starting point is not a case; it is a mood.
A named person owns the outcome
Not a function, not a steering group, a person. When the benefit is everyone's, it is no one's. The question "who loses sleep if this delivers nothing?" should have a one-word answer, and that person should be in the room.
The assumptions are written down as things that can fail
Every case rests on assumptions: adoption rates, data quality, a supplier's roadmap, a regulation staying put. Weak cases bury them in an appendix. Strong ones list them, state what happens to the benefit if each one breaks, and say who is watching each one. An assumption you have named is a risk you are managing. An assumption you have hidden is a surprise you have scheduled.
The case says what "off track" means before anything is off track
Decide, at the moment of approval, what deviation looks like and what happens when it appears. If you wait until the numbers disappoint to define disappointment, the definition will be negotiated by whoever is most embarrassed. The organisations that do this well treat a missed target as a trigger for a recorded decision: revisit, rescope, restate or stop. The ones that do it badly treat it as a communications problem.
Failure has somewhere to go
Some investments will not deliver. A credible case admits this at the start and defines what will be learned and kept if it happens. Boards do not punish honest failure nearly as often as people fear. They punish surprise, and they remember being managed.
None of this requires sophistication. It requires the willingness to be held to something. And that is exactly why vague cases exist: a number nobody can trace is a number nobody can miss.
The uncomfortable conclusion is that the first committee meeting is not where good AI cases go to die. It is where untestable ones do, and that is the committee doing its job. If your case cannot survive the questions "where did this number come from", "who owns it" and "what happens when it is wrong", the kindest thing the committee can do is kill it before it costs anything.
Build the case to be tested. Then the meeting stops being the obstacle and starts being the evidence.