Put a number on your AI spend that finance can actually defend

What makes an AI spend number defensible

Key takeaway: finance can defend a number when it can show where the number came from, explain why it changed, and say roughly what it will be next quarter.

A large AI bill is manageable if finance can explain it. The trouble starts when the only thing anyone can say about it is the total.

Finance tends to test any spend line in three ways. Can it be traced to the work that caused it? Can someone explain why it went up or down? Can it be forecast with reasonable confidence? A cloud bill with good tagging passes all three. AI agent spend often fails the first one, and the other two depend on it.

Practitioners see the same gap. In the FinOps Foundation's State of FinOps 2026, they named visibility into AI costs as their top challenge in managing AI spend, followed by allocating those costs to business units. The tool capability they most want and can't yet get is granular monitoring of AI spend: tokens, model requests and GPU use.

The five levels, and what each one answers

Key takeaway: each level of the cost breakdown answers a different question, and a number finance can defend needs all of them working together.

Two of these levels do most of the work.

The workflow level is the one finance gets the most out of. A workflow is a repeatable kind of agent work, such as reviewing a pull request or turning a ticket into code. Once you know the average cost of running a workflow once, its unit cost, you can multiply it by how often it runs. That turns AI spend into something that looks like any other usage-based budget line.

For engineering, the run level matters most. A run is a single piece of work: one pull request reviewed, one ticket picked up. When a number looks wrong, the run is where the explanation lives.

How finance builds the number

Key takeaway: start from the total, split it by team, then use cost per workflow and expected volume to forecast.

1. Start with the organization total for the period. This is the number the CFO already has.
2. Split it by team. That shows who owns which part of the budget. It also supports showback, where each team is shown what its usage cost without being billed for it.
3. Within each team, look at workflows. For each one, note the average cost per run and how many runs there were.
4. Forecast from unit cost and volume: multiply each workflow's cost per run by the number of runs you expect next period.

To put some illustrative numbers on it, say a team's pull request review workflow ran 2,000 times last month at an average of $0.40 a run, for a total of $800. The team expects to review about 3,000 pull requests next month, so the starting forecast is $1,200.

That forecast is a midpoint. Individual runs vary, sometimes by a lot, for reasons we covered in $0.50 yesterday, $5 today. Look at how much the monthly average has moved over the last few months and put a range around the midpoint based on that. A range finance can see the reasoning for is easier to defend than a single figure that turns out to be wrong.

How engineering explains the outliers

Key takeaway: when a number looks wrong, engineering starts at the run level and works up.

The workflow average gives you a baseline. A run that cost far more than that baseline is the one to open. From there, engineering can see which model it used and what the agent did along the way. That level of detail is still rare. KPMG's Q2 2026 AI Pulse survey of large US companies found that two-thirds have AI monitoring dashboards, but only 26% have full, real-time visibility into what their AI systems cost to run.

The usual causes are familiar: the agent retried several times, its context grew over a long run, or it read far more of the codebase than the task seemed to need. Sometimes the fix is a change to the workflow, such as tighter instructions or a different model for that step. Sometimes the task was simply harder than it looked, and the cost was reasonable.

Finance can accept either answer, as long as someone can give one.

Doing this in Upsun Dispatch

Key takeaway: Upsun Dispatch records cost by run, workflow and model as work happens, so most of the breakdown exists without anyone having to build it.

Upsun Dispatch™ records the agent time each run used and an estimate of its token cost, and the Usage page breaks spend down by workflow and model. To see spend by team, give each team its own Dispatch organization; they can all draw on the same billing profile.

Each run's final cost is logged alongside everything else the run did, so engineering can go straight from an outlier to an explanation. Token costs are estimates based on your AI provider's published prices, so finance should check the final figure against the provider's invoice.

Agent time is drawn from credit you fund in advance, which sets the ceiling for that part of the bill. Model tokens are billed by your AI provider. We cover how that works, including what happens when the balance runs out, in our guide to spend limits.

It has limits worth planning around.Only agent time is capped by your credit, and there are no separate caps for each team or workflow, so those budgets are tracked against the breakdown rather than enforced.

The breakdown also only tells you what happened and where. Deciding what a workflow should cost is still a conversation between finance and engineering, though it's a much shorter one when both sides are looking at the same numbers.

See the cost view in action. Follow one code review from top-up to the Usage page.

Frequently asked questions (FAQ)

What is cost per workflow?
It's the average cost of running a workflow once, for example the average cost of one AI pull request review. Multiplied by how often the workflow runs, it gives you a forecast for that kind of work.

What's the difference between showback and chargeback?
Showback means showing each team what its usage cost. Chargeback goes a step further and bills that cost to the team's own budget. Both depend on spend being broken down by team.

How do you forecast AI agent spend?
Take each workflow's average cost per run, multiply it by the number of runs you expect, and add the workflows up. Because individual runs vary, present the result as a range based on how much the average has moved in recent months.

Can Upsun Dispatch set a separate budget for each team?
Not as a separate cap. Agent time is capped by your credit as a whole. You can give each team its own Dispatch organization so its spend is tracked separately, but organizations sharing a billing profile draw on the same balance, so team budgets are tracked rather than enforced.