The list is never the problem
Most companies have more AI ideas than capacity. The difficulty is that the ideas are ranked by how impressive they sound in a meeting, which correlates weakly with what they return and strongly with what they cost.
Axis 1: how expensive is the workflow today?
Score by the recurring cost of the work being replaced, in hours or headcount. If nobody can state that number, the idea is not ready — start at the most expensive recurring workflow, not the most visible one.
Axis 2: can a human verify the output quickly?
Features where correctness is checkable in seconds are safe to ship early and generate correction data immediately. Where nobody can tell right from plausible, the feature should wait.
Rank by cost of the work and cheapness of the mistake. Excitement is not an axis.
Axis 3: what does a mistake cost?
Recoverable, invisible, or reputational. Anything customer-facing or irreversible carries governance work that multiplies the effort, which is why internal drafting beats autonomous customer conversation as a starting point.
Axis 4: does the data exist yet?
Some ideas require documentation or decision history nobody has. That is not a blocker but it is the real scope, and it is why the knowledge base often is the project.
Then sequence for compounding
Prefer the feature whose by-products help the next one — a retrieval layer, a corrections dataset, an evaluation harness. Three features sharing infrastructure cost far less than three built independently, and the first one should be chosen partly for what it leaves behind.