"So what did we get for it?"
Your CFO asks some version of that question about last quarter's AI spend. Pulling the bill takes four minutes. Answering the question takes a project nobody has staffed.
A director of AI infrastructure at a large university put it plainly to us on a call, saying, "The bottom line is it's a mess because I have no idea where the money is going or what's the return on that money."
We went back through every buyer conversation we had in the first half of 2026. At 41 of the 82 accounts we spoke with, the same issue came up without us asking.
It came up more often than build versus buy, bill shock, or any competitor comparison. Nearly all of them could tell us what they were spending, at least at the invoice level.
Almost none could tell us what they got in return.
Everyone can already see the cost
The spend itself is visible. In planning meetings at a very large retailer, the list of things the company wants to build keeps growing, and nobody ever raises what any of it returns. Their FinOps lead pressed colleagues for the business case behind one of those plans and got a shrug. "I don't really know. I just know I need to do it." Those shrugs get expensive fast.
There is an investor version of the same conversation. What did $100,000 of AI spend produce? A CTO at a workforce analytics company told us the honest answer is that they haven't been tracking those metrics.
Work that stalls out gets a name at a global consultancy. Their product director calls it a science experiment, because the team can't articulate the value to anyone outside of engineering.
Company size doesn't protect you either. One global professional network spending well over half a million dollars a year on AI still couldn't connect developer tooling spend to developer output, and their strategic finance lead told us they weren't yet seeing the productivity the spend implied.
Why your dashboards cannot close it
There were three structural reasons that showed up again and again.
The unit of measurement is wrong. A FinOps lead at a major cloud provider warned us against anchoring on cost per token, because that metric does not tell the whole story. Tokens measure consumption, which sits a step removed from anything a budget owner can approve or defend.
The trail runs cold. AI costs can't be traced back to the people responsible for them, which is what customers keep reporting to one cloud cost management vendor we spoke with. Provider bills roll up by key and by account, while work happens by ticket, customer, and agent. Nothing in the default tracking joins those two things together.
The denominator is missing. An engineering leader at a large Google partner made the point in one image, describing a model that "burned freaking $10 answering your question, only to get it out wrong. And you're like, well, what did I pay for right now?" The charge was accurate, but it had no outcome attached to it.
Forecasting inherits all three problems at once. A cloud and AI cost lead at a security company was refreshingly blunt about where that leaves everyone. "I'd be lying if I said we figured out forecasting. And I think anyone who says they figured out forecasting is lying. Literally everybody."
What the teams making progress do
In reality, very few have this solved, and not because nobody thought of it. Attribution is engineering work with no obvious owner. Someone has to tag calls where they are made, that work competes with shipping features, and it produces nothing visible until enough data accumulates to answer a question. The cost is immediate and the payoff is a quarter away.
The ones moving fastest start by attaching spend to something the business already recognizes, which usually means a ticket, a support case, a customer account, or a named agent instead of an API key. They instrument at the call site, so they record the attribution as the money goes out rather than reconstructing it from a bill three weeks later. Once cost lands on a familiar object, finance can reason about it without a translation layer.
They also settle the manual baseline before anyone claims a saving. Most ROI conversations skip a prerequisite, which is an agreed number for how long the work took before any AI touched it. A partnerships lead at a payroll and HR platform named it for us. Seventeen minutes to prepare a sales proposal, in their case.
That one number does a lot of work. Multiply it by proposals per month, price the minutes at a loaded rate, and you have a figure to set the AI cost against. Without it, every efficiency claim is an argument. With it, the math is arithmetic.
And they judge the result by yield. A multi-agent system for support cases at another security company would run roughly $400,000 a year at full volume, and the enterprise IT lead who built it framed the decision the way a CFO would. "Is that worth the yield that you're getting from deflecting these support cases?" A deflection rate measured against a cost per case gets you an answer to that.
The furthest along was a network security company where the board now asks for token usage by engineer, mirrored against merged pull requests. Whatever you make of that as a productivity metric, it gives them a denominator and a review cadence, which is more than most companies have.
Where the cost should land
The surprise is never the total. In a first month of attributed data, what surfaces is the single misconfigured key that ran up roughly $90,000 in a week, the agent running open loop with no sense of when it should stop, or the expensive model quietly handling a job a cheaper one could do. None of those show up on a monthly invoice. All of them show up the moment cost is attached to the work that caused it.
That is what we built Revenium to do. You capture usage at the call site through the SDK and middleware and tag it with the business context that matters to you, so the cost lands on a customer, an agent, or a ticket from the start.
From there, you can analyze decision costs by agentic job rather than by individual call, record outcomes when a job actually produces something, and read ROI and unit economics against those outcomes instead of against raw consumption.
You will get asked the ROI question again next quarter. Decide now which business object you want the answer attached to.
