What Boards Should Ask Before Approving an AI Budget

AI spending has moved from innovation budgets into operating budgets at most large organisations, and the numbers are now material enough that boards are asking harder questions.

The difficulty is that the standard questions do not work well here. Return on investment estimates for AI programmes are frequently unfalsifiable. Here are five that produce more useful answers.

What decision or process does this change?

A proposal that describes a capability rather than a change is not yet an investment case. Ask which specific process will operate differently, who currently performs it, and what they will do instead.

If the answer is that the technology will make people generally more productive, the benefit will not be measurable and will not be realised.

What happens when it is wrong?

Every one of these systems will produce incorrect output. The question is what the consequence is and who catches it.

A proposal that has not thought about the error case is not ready. This also usefully separates low-risk internal applications from customer-facing ones, which frequently arrive in the same budget line with the same governance.

What is the total cost, including the parts nobody quotes?

Licence cost is the visible number. Data preparation, integration, ongoing evaluation, monitoring, human review capacity and the internal time to manage the vendor usually exceed it.

Ask for the fully loaded figure over three years. The ratio between that and the licence line is informative about how carefully the proposal was prepared.

Could we stop?

Ask what happens if the vendor doubles the price, changes the model, or ceases the service. Where the answer is that the process could not run without it, you are accepting a dependency that belongs in the risk register.

This is becoming a live commercial question as pricing in the sector continues to move.

How will we know it worked?

Agree the measure before approval, not afterwards. It should be a business metric that existed before the project: cycle time, error rate, cost per transaction, conversion.

Adoption statistics and usage counts are not outcomes. A programme measured on how many people logged in will report success indefinitely without changing anything, and the board will have no basis on which to stop funding it.

It is also worth agreeing in advance what result would cause you to stop. A proposal with no failure condition is not an investment case, and boards that ask this question early tend to receive better proposals the following year.