A $200-per-seat license looks predictable until you multiply it by everyone, including the people who ask the AI three questions a month. Raw API usage looks cheap until one workflow loops all weekend. Both bills are avoidable with one spreadsheet built on your actual workloads.
We model it per team: heavy creators on seats, sporadic users on metered APIs, automated workflows on capped service accounts. The blended number is almost always lower than either default.
The artifacts that earn their keep
A per-workflow cost map that a finance analyst who was not in the room can correct. If they cannot, it is a poster.
A short list of spend decisions that are expensive to reverse, each with the option you are not taking and why.
One capped pilot, metered end to end, before any renewal. A model that has not survived contact with the real bill is still a hypothesis.
What to leave out
Vendor comparisons without your workloads. Target-state posters with no migration. Projections that require adoption you have not measured. If the plan needs everyone to become a power user to pay back, say that in the first paragraph.