Why most boards still can't answer what AI actually costs to run

I saw a stat this week that made me sit up, not because it is shocking, but because of exactly how unshocking it is once it has thirty seconds to sink in. KPMG surveyed 204 C-suite and senior business leaders at companies with a billion dollars or more in revenue, fielded between late April and late May, and published the results on June 24. Only 26% of them said they have full, real-time visibility into what their AI systems cost to operate. Two thirds already have monitoring dashboards running. Two out of three still cannot answer the question their own dashboard was supposed to answer.

Why dashboards are not the same as an answer

Here is the part that will land on a board’s radar, if it has not already: this is not a story about companies that ignored AI cost tracking. Sixty six percent have dashboards. Sixty one percent have approval processes for AI spend. Thirty six percent have implemented direct token or usage controls. The tooling exists. The number still is not there.

CFO Dive’s coverage of the KPMG survey put it plainly.

The reason I think this matters more than the usual budget-overrun story is what is happening on the other side of the ledger at the same time. The same survey found that coordinated multi-agent orchestration, meaning agents that work across functions rather than one bot doing one task, doubled from 9% to 18% of organizations in a single quarter. Complexity is compounding faster than the ability to price it. That is the actual gap. Not “AI is expensive.” Everyone already knows AI is expensive. The gap is that the systems are getting more entangled while the meter stays fuzzy.

There is an older line I keep coming back to on this exact point, from a FinOps Foundation report published earlier this year and revisited at their June conference: “Is your AI providing value? No one can answer that question yet.” That was written before the multi-agent doubling KPMG just measured. It has aged, if anything, more true, not less.

What your board will ask about the AI bill nobody can fully explain

Do we actually know the number, or do we know we have a dashboard? These are different claims. Sixty six percent of companies in the KPMG sample have dashboards. Only 26% have the real-time answer. If your CFO’s confidence comes from “we bought the tool,” ask a follow-up: what was the exact AI cost last month, broken out by workflow, and how fresh is that number.

Are we ahead of our industry or behind it? This is the part I found genuinely useful, because visibility is not evenly distributed. Technology companies in the KPMG sample reported 54% full visibility. Banking came in at 31%. Asset management and private equity firms reported just 4%.

That is not a small gap. If your company sits in one of the lagging sectors, “everyone struggles with this” is technically true and also not the reassurance it sounds like.

As we add more coordinated agents, does our visibility keep pace, or fall further behind? Multi-agent orchestration doubling in one quarter is a scaling event. A separate practitioner survey from LangChain, covering more than 1,300 engineers, found 89% have implemented some form of agent observability already, and quality and latency, not cost, are still the top blockers to production. Instrumentation is not the bottleneck. Turning instrumentation into a number the board can trust is.

The sixty-second version for your board

Say the number honestly. A company inside the 26% should say so and show the workflow-level breakdown. A company outside it should say that plainly too: “We have monitoring in place, we do not yet have a real-time cost answer, and here is the three month plan to get one.” That sentence, said early, is worth more than a confident guess said late. Boards forgive a gap with a plan. They do not forgive discovering the gap themselves, from an invoice.

Buying the dashboard was the easy step. Turning it into an honest number was always the actual job.

What determines whether this gap closes or widens

I do not think this closes on its own, and I do not think it needs a heroic effort either. The companies that will close it are the ones that treat cost attribution as an operating discipline with a named owner, the same way they would treat revenue recognition, rather than a side effect of whichever monitoring tool procurement bought last year. The ones that will not close it are the ones still waiting for the dashboard to do the work the org chart was supposed to do. Watch the multi-agent count over the next two quarters. If it is climbing and the cost-visibility number is not climbing with it, that is the conversation to have before your board has it for you.