The question your CFO keeps asking, and why you cannot answer it yet

The meeting where the question comes up

Key takeaway: finance's questions about AI spend are easy to ask and hard to answer, because the data needed to answer them was never collected.

It's the quarterly budget review. The AI line has grown again, and the CFO wants to know what it paid for. Which teams are driving it? Is it going to keep climbing? What would it cost to give agents twice as much work next quarter?

The CTO has a total from the model provider and a fair sense of which teams use agents most. Pressed for detail, the best the CTO can offer is something like "roughly, I think." That's a hard number to put in a budget.

These are fair questions. Finance asks them about every major spend line, and for cloud in particular it has come to expect a clean breakdown. In the Flexera 2026 State of ITAM Report, 78% of organizations said they now have a dedicated FinOps team. AI spend is being held to the same standard, and plenty of teams can't meet it yet.

Why the answer isn't there yet

Key takeaway: AI tools tend to arrive one team at a time, and their cost reporting shows who used them, while finance wants to know what the money was spent on.

In many organizations, AI agents didn't arrive in one planned rollout. A team tried a tool, it worked, another team signed up, and a few months later there's a handful of accounts and keys, plus a bill that has quietly become significant.

Out of the box, model providers typically report usage by account, project or API key. That shows who holds the key, but not which ticket, pull request or workflow the tokens went on, and that's the level finance is asking about.

Agents make the gap wider. A single workflow might call more than one model, retry when an answer falls short, and read far more of the codebase than the task seemed to need. (We covered why that makes individual runs so unpredictable in $0.50 yesterday, $5 today.) Even a team that tracks its provider usage carefully ends up with a number that's accurate and hard to explain.

The Flexera survey points the same way. Only 31% of organizations said they had accurate visibility into the AI software in their IT estate, the lowest of any category in the survey. If teams aren't sure what AI tools they're running, a per-workflow cost breakdown is a long way off.

Make cost transparency part of the buying decision

Key takeaway: the cheapest time to make AI spend answerable is before a tool is rolled out, by making cost reporting part of the evaluation.

Once a tool is embedded across several teams, adding cost attribution after the fact is slow work. Someone ends up tagging keys and building spreadsheets, then chasing teams for context on last month's spikes. Asking the right questions during evaluation avoids most of that.

Here's what to ask any vendor whose product runs AI agents on your behalf:

• Can we see cost per run and per workflow, or only per account?
• Is cost visible while work is running, or only after it finishes?
• Can spend be broken down by team, and by the model and provider used?
• How do we cap spend, and what happens to work already in progress when we reach the cap?
• Can we set separate budgets for each team or workflow?
• Are we billed while work sits waiting for a person to approve it?
• Is the final cost recorded alongside a record of what the agent actually did?

Listen for answers that start with "you can export the data and…". Those usually mean the CFO's question will land back on engineering's desk, just a few steps later.

What an answerable AI bill looks like

Key takeaway: when every unit of spend is tied to the work that caused it, finance's questions can be answered from a report.

Upsun Dispatch™ answers most of that list. Every run records its agent time and an estimate of its token cost, and the Usage page breaks spend down by workflow and model. Teams that need their own numbers can each have their own Dispatch organization, all drawing on the same billing profile. Engineering can open an individual run to explain an outlier.

Agent time comes out of credit you fund in advance, in USD, CAD, GBP, EUR or AUD, and you're notified before the balance runs low. Model tokens are billed by your AI provider directly, with no markup from Upsun.

When it runs out, no new runs start until you top up, and runs already in progress are allowed to finish. Agent time is only billed while a sandbox is running, so a run paused at a gate, waiting for someone's approval, doesn't use agent time. When a run finishes, its final cost is logged against it, alongside everything else it did.

Two questions on that list get a partial answer. Your credit caps agent time, but token spend sits on your AI provider's bill, so Upsun Dispatch tracks it per run rather than capping it. And there are no separate caps for each team or workflow, so those budgets are tracked against the breakdown rather than enforced.

Next time the CFO asks, have the answer ready. Learn how to break AI spend down by team, workflow and run

Frequently asked questions (FAQ)

What does a CFO need to know about AI agent spend?
Usually four things: how much is being spent, which teams and workflows are driving it, how it's trending, and what it would cost to do more. A provider total answers the first. The other three need spend tied to the work that caused it.

Why can't model provider dashboards answer finance's questions?
They typically report usage by account, project or API key. That shows who used the model, but not which workflow, ticket or pull request the spend went on, which is the level finance plans and approves at.

What should you ask an AI vendor about cost before buying?
Ask whether cost is visible per run and per workflow, whether it's visible while work is running, how spend can be capped, what happens to work in progress at the cap, and whether you're billed while work waits for human approval.

Does Upsun Dispatch bill while a run waits for approval?
While a run waits at a gate, the sandbox is shut down, and a new one starts once the run is approved.