Before the AI pilot, define the operating decision.

Start with a decision, not a tool.

Describe who needs to act, what they need to know, and why the current process is insufficient. This makes it easier to distinguish a valuable business problem from an interesting technical demonstration.

Examine the dependencies.

Data availability is only one consideration. The process may need a new approval step, an integration with an existing system, or a different allocation of responsibility. Document these dependencies alongside the model requirements.

Define evidence for the next investment.

Agree what a pilot must demonstrate before it scales. That may include output quality, review effort, adoption by the intended users, and the effect on the underlying workflow. A successful demonstration alone is not the same as an operating capability.