Build a cost model you can recalculate

I'd build the model out of variables instead of one annual figure. Keep price assumptions, observed rates and policy decisions in separate fields, so when one of them changes you're updating a cell and not rebuilding the whole worksheet.

For each unit, estimate or measure:

• average and high-percentile input size
• model calls per stage, including retries and evaluation calls
• input, cached input and output usage, where the supplier prices them differently
• retrieval, storage, queue and compute usage
• third-party extraction, search or verification charges
• expected human handling time, by role
• the share of units that go down each exception path
• fixed monthly platform and observability costs
• engineering, operations, security and content maintenance time

Run a base case and then a few volume and failure scenarios. I'd be careful about calling any of them "worst case" unless you can actually defend it as a bound. A useful stress case might combine a volume peak with a provider incident and a higher manual review rate. Put every assumption right next to the result it feeds.

Keep variable costs apart from step costs. A queue worker might absorb more volume until it hits a threshold, and a compliance review or support roster might add a whole block of work when the workflow expands into another business unit. The model should show those jumps.

Write down the currency, the tax treatment and the date of each price. Supplier prices and exchange rates move. Record where each rate came from and who's going to refresh it before a budget decision gets made. If the contract has volume tiers, minimum spend or data residency charges, model those terms properly instead of multiplying a public unit price by your volume.

And don't hide free allowances. They're handy during a trial, but you can't build steady-state economics on them. Show the cost before and after credits, and note when the allowance runs out.