Why AI cases fail approval
Not because the technology is doubted. Because the case describes a capability rather than a saving, and finance cannot approve a capability.
The fix is arithmetic that someone can check, on one page.
The five parts
- Current cost — hours, rate, volume, error cost, with how you measured it
- Proposed cost — build, running, maintenance, over three years
- The difference, with the payback period
- Risks, each with what you would do about it
- The measure — one number reviewed at ninety days
Anything you cannot measure yet becomes a stated assumption with a source. An honest assumption is far more persuasive than a confident number nobody believes.
Be conservative on the saving
Assume the lower end of the accuracy range, the higher end of the cost, and that not all saved time is recovered. If it still clears, the case is strong.
Optimistic cases that miss are what make the next three projects harder to approve.
Include the risks properly
| Risk | Response |
|---|---|
| Accuracy below target | Proof of concept first, with a stop point |
| Team does not adopt it | Volunteers first, involved from week one |
| Running cost higher than modelled | Hard cap with an alert |
| Supplier dependency | Code and credentials owned by us |
| Process changes | Phase one only, reassess after |
A case that names its risks reads as considered. One that names none reads as unexamined.
Ask for the small thing first
A proof of concept at a few thousand pounds is a much easier approval than a full build, and it makes the second approval nearly automatic.
It also protects everyone if the answer turns out to be no.