Budgeting and governing an AI workflow after the prototype

A prototype budget usually lists the obvious ingredients: a model, some development time and maybe a data store. Once the thing is running every day, it picks up work the demo never had to carry. Inputs turn up in bad shape, permissions change, someone has to review the uncertain results, failed runs need cleaning up, and when a user disputes something the workflow did, somebody has to answer them.

So I'd build the cost model around the whole workflow. You still need volumes, service prices and infrastructure, but you also need staff time, how often things go to exceptions, maintenance, controls and what it costs to fix a bad outcome. Governance is then mostly a matter of giving each of those pieces an owner, deciding how often it gets looked at and who makes the call when something needs deciding.

This is a way of planning, so don't read it as a benchmark. Prices, labour rates and how much risk an organisation will wear all vary. When you fill it in, use current supplier terms and whatever workflow data you've actually observed, and keep your assumptions in a different place from your measured results.