The line item that is easiest to defend is rarely the real cost
When we moved petabytes of data from on-premises data centers to a cloud data platform and modernized the enterprise data warehouse, compute and storage were the easiest line items to defend in front of the CFO. They were also the least accurate predictor of total cost. The real effort lived in validation, reconciliation, access reviews, regulatory reporting sign-off and the people who had to trust the new numbers.
Enterprise AI budgets repeat that mistake. Model and inference pricing gets the scrutiny, while data preparation, permission review, reference answers and user support hide inside a general transformation budget. My starting point is the complete workflow, including the work people do around the model.
Price the workflow, not the query
Map the current workflow before estimating savings. Identify who gathers records, reconciles disagreements, prepares the response and approves the decision. Distinguish elapsed time from actual effort. A report delivered after a long queue may involve little analyst work, while a quick executive answer may tie up several people at once.
Choose the unit of value carefully. Cost per generated answer helps with infrastructure planning, but cost per accepted decision is what a sponsor should fund. A cheap answer that needs heavy correction is expensive in practice. A pricier request can be worth it if it produces inspectable evidence and removes rework.
I learned this in regulatory reporting. When a data warehouse feeds numbers that an examiner or audit committee will rely on, the cost of a wrong figure is not the rework. It is the loss of trust in every figure that follows. An AI answer that informs a credit, pricing or clinical operations decision carries the same exposure, so the budget should pay for the evidence that keeps that trust intact.
Fund the one-time work honestly
Initial costs include connecting sources, documenting definitions, configuring access and assembling evaluation examples. Some of that improves the broader data environment even if the pilot stops. Keep that value visible without crediting all of it to AI, or the initiative will claim benefits that a separate data governance effort created.
Budget business reviewer time explicitly. A compliance officer's or clinical reviewer's hours are not free because they do not generate an invoice. In banking, healthcare and life sciences these reviewers are usually the binding constraint, so the evaluation schedule should reflect their real availability.
Do not treat every prerequisite as a permanent blocker. A pilot can run on a narrower source or smaller question set while larger remediation proceeds. Narrow scope is useful when it reduces uncertainty. It is misleading when presented as proof of enterprise readiness.
Plan for the costs that recur
Recurring costs include inference, hosting, connector maintenance, monitoring, support, reevaluation and onboarding. Estimate a range from expected volume and concurrency, then compare it with actual pilot behavior. Never borrow a cost per query from an unrelated task.
Count unsuccessful work. Retries, timeouts and clarifications consume resources and attention. A high retry rate from an unstable source calls for a different investment than a high clarification rate from vague questions.
Factor in change frequency. A workflow built on stable monthly reports needs little maintenance. One spanning fast-changing applications needs continuous attention. The same integration price can produce very different operating costs.
Fund decisions in stages, sized to the institution
Divide the budget into stages that each resolve one uncertainty: access and reference examples first, answer quality next, support capacity with more users after that. State what evidence justifies each increment and who decides.
Use conservative benefit assumptions until you have observed the workflow. Multiplying every saved minute by a salary rate and calling it cash savings will not survive a CFO's review. Some benefits are capacity, speed or less rework. Describe them accurately and ask how the released capacity will be used.
Check the budget against the people named in the delivery plan. A small team may be assigning the same person to source cleanup, reference answers and user support. Those activities compete for time even when no additional salary appears in the estimate. Where independent validation or third-party review is required, include it in the schedule and cost from the start.
A credible budget makes tradeoffs discussable: the cost of learning, the cost of operating and the conditions for expansion. That is far more useful than an optimistic return built on the price of model calls alone.