An operations leader asked to fund an AI assistant needs to know what staff will do differently when it arrives. A demonstration may show that the model can read emails, compare documents and prepare a convincing answer. It does not show whether that answer will reach the right person before a commitment is made, or whether checking it will cost more effort than preparing it by hand.
The AI business case starts with that work. For a fast-moving consumer goods (FMCG) operation, the records may sit across commercial, packaging and production teams. The investment question is whether bringing them together changes a decision enough to justify the cost. Model capability matters once the decision and its constraints are clear.
Show how the improvement would pay for itself
Starting with what a model can do encourages a weak calculation: estimate how much faster it completes a task, multiply by staff cost and call the result a saving. That leaves out the people who must check its answer and the work needed to act on it. Available staff time becomes a cash saving only if spending falls; it may instead provide capacity for other work, whose value needs its own explanation.
Industry survey evidence from Deloitte and the joint BCG/Consumer Goods Forum research in 2026 points to a gap between AI ambition and return measurement in retail and consumer goods. These are executives' reported priorities and measurement practices, not audited returns. They establish a reason to examine the investment case carefully, not evidence that a particular assistant will pay for itself.
The missing link is how an output changes an operating choice. A more complete answer has little commercial value if the retailer has already received a promise that the plant cannot fulfil. Accuracy after the decision deadline will not repair that sequence.
Describe the decision before choosing the tool
Set a boundary around the proposed change. Name who prepares the evidence, who can make the commitment and when they need to decide. Trace the records they use today, including how they resolve conflicting versions. The baseline should include waiting, checking and correcting mistakes, rather than just the time spent writing the final note.
Cost the change people must adopt
In this example, assembling a note is only part of the investment. Someone must give the assistant access to approved records, identify obsolete specifications and keep supplier confirmations current. Staff need time to learn the process and challenge unsupported answers. Include that adoption effort, connections to existing systems and ongoing review alongside purchase and running costs, over the same period as the benefits.
The assumption most likely to reverse the choice is whether usable confirmations exist before the commitment deadline. If staff still have to obtain them by phone, reading emails faster may change very little. Test that dependency before funding a wider deployment. Where the simpler approval record solves the problem at lower cost, it should win the comparison.
| Required evidence | Question it answers |
|---|---|
| The decision and its deadline | Who must decide what, and by when? |
| Current work, including waiting, checking and rework | What effort and delay could the change reduce? |
| The adoption cost | What will access, training and ongoing review cost? |
| The assumption that could reverse the choice | Which dependency could remove the expected benefit? |
| The evidence worth buying first | What should we test before committing more money? |
Let the evidence determine the outcome
Approve when the evidence supports a worthwhile change after adoption and checking costs. Defer when a specific unanswered question could change that judgement, such as whether confirmations can arrive in time. Reject when the benefit depends on unavailable records or a simpler change does enough. None of these outcomes is a failure of AI evaluation.
Before asking which model to buy, ask the operating owner to show the current decision record and the point at which better evidence would change the commitment. If that point cannot be identified, a model demonstration is premature. The work gives the business case its boundary and gives approval something concrete to depend on.