Finance asks for next quarter's AI budget. Platform has a working pilot, three teams waiting to join, and no stable estimate of how often those teams will use it. A single number looks decisive and hides two questions: how much demand will arrive, and how much the company wants to fund.

Use a range with release conditions. The operating budget describes a workload, an accepted unit of output, and a decision that releases the next tranche. It also separates an internal spending decision from the provider account's ability to process requests.

A finance practitioner asks how to pressure-test an AI spending forecast when the estimate comes from IT. Start there: identify assumptions the budget owner can examine, then separate forecast demand from authorized spending.

Put volume and unit cost on different lines#

Begin with accepted outputs and skip active users. A team of five whose agent processes a daily queue can create more work than a team of fifty. Seats are a purchasing unit and predict neither model usage nor review demand.

Estimate demand in three bands: already committed work, likely adoption, and optional experiments. Estimate unit cost separately for ordinary tasks, difficult tasks, and failed attempts. Keep a reserve for investigation and recovery outside the cost of normal delivery, or every incident shows up as unexplained overrun.

Take a hypothetical quarter with 1,000 ordinary tasks at $2 each, 200 difficult tasks at $8 each, and a $400 experiment allowance. Variable spend is $4,000. A higher-demand case with 1,500 ordinary tasks and 400 difficult tasks reaches $6,600 before any change in unit cost. These are planning inputs, not vendor prices.

Stress unit cost independently. If difficult tasks take twice the review or model work, the same high-demand case grows again, and finance can see whether the main uncertainty is adoption, complexity, or both.

Write the release conditions#

Funding band	What it pays for	Evidence before expansion
Committed operation	Known recurring work	Named owner and accepted output
Adoption reserve	Additional teams or volume	Measured unit cost and available reviewers
Experiment allowance	New task types	Fixed end date and spending boundary
Recovery reserve	Investigation and safe completion	Incident owner and documented use

Release conditions should be observable. "Promising adoption" is hard to challenge. "The new team completed thirty representative tasks with recorded review time and a tested stop procedure" tells the sponsor what must happen.

The upper range is not permission to spend automatically. Name the person who can release each band and the activity it covers. If a team shifts from summarizing records to taking external actions, revisit the scope even when its cost stays inside the original range.

Plan the week that exceeds the range#

Have finance approve in advance what gets reduced first. Optional experiments, duplicate refreshes, and low-priority backlog are deferrable. A customer commitment due today needs a manual fallback.

Separate a service interruption from a funding decision. If a provider rejects requests while the internal budget remains available, raising the internal budget will not repair an expired credential or account restriction. Use the troubleshooting questions to find the failing point.

If demand arrives faster than expected, do not multiply yesterday's average cost by every queued task. The next cohort can contain larger accounts or less standardized inputs. Sample it, record the difference, and update the forecast.

Reforecast around decisions#

The FinOps forecasting guidance connects forecasting with budget decisions. A useful monthly review explains variance as volume, mix, unit cost, or fixed cost. More completed work at the same unit cost is a different situation from the same work requiring more attempts. Both increase the invoice, and only the first clearly indicates growing demand.

Bring agent spend, accepted-output counts, and reviewer capacity to the same meeting. Use the budget ownership guide to settle who can authorize expansion across teams. The deliverable is a revised range and a specific next funding decision.