Almost every organisation we work with has run an AI pilot. Far fewer have taken one to production at a scale that shows up in the accounts. The blocker is rarely the technology and rarely the appetite. It is that the business case arrives in a language the finance function cannot underwrite.
A pilot report that says the model summarises claims correctly in 91 per cent of cases is a capability statement. A CFO reads it and asks a different question: what changes in the cost base, and when?
Three things the case has to answer
- Whose time is released, and what happens to it. Two hours saved per person per week is not a saving until either headcount changes or that time is redirected to work you would otherwise have hired for. Say which one it is.
- What it costs at full volume. Inference costs at pilot scale mislead. Model the cost at the volume you actually process, including the reprocessing that error handling will generate.
- What breaks when the model is wrong. Quantify the cost of an error and the cost of the review step that catches it. A process that needs a human check on every output has a much smaller benefit than the pilot suggested, and that is fine as long as it is in the number.
Size the process, not the technology
The most useful hour in an AI business case is usually spent away from the technology entirely, measuring how the process runs today: volumes, handling times, rework rates, and the proportion of cases that are genuinely standard. In one recent engagement that exercise showed only 38 per cent of the transaction volume was standard enough to automate. The case still worked, but at a third of the original claim, and it survived review because the number was defensible.
A smaller benefit that finance believes will get funded. A larger one that they cannot verify will not.
Stage the funding
Ask for the amount needed to reach the next decision point, not the full programme. A first tranche that covers one process end to end, with agreed success measures and a stop date, is far easier to approve than a three-year platform commitment. It also gives you real production numbers to build the second case on, which is the only thing that makes the second case easy.