Coordination FinOps: The Category No One Has Measured
Cloud FinOps measures infrastructure. AI FinOps measures inference. Human-agent coordination is the missing third layer, and it carries the bulk of the real cost of operating AI in the enterprise.
90-Second Summary
FinOps exists because one day the cloud bill came in too big and nobody could explain where it came from. The discipline gave that spend a name, an owner, and a number. Then it stretched to cover AI inference, tokens and GPU, the cost the machine burns while it runs as technology. Solved. What is still missing is the third reading, and it is the one that weighs the most: what the machine charges your humans to close a decision. Model a 500-person SaaS at average adoption and AI infrastructure fits inside a small slice of the total cost of the hybrid operation. The rest, the larger part, lives in the coordination edges, diluted into payroll with no category of its own. That third reading has no dashboard, no closed playbook, and shows up in no vendor RFP. It is exactly where the biggest slice is draining away.
Open the spreadsheet for your next budget review. Cloud spend is broken out by workload, by team, by quarter, clean to look at. The AI inference line showed up last year and grows at the pace of something young. Decent visibility on both. Then you pull the payroll report. Your senior headcount is flat, same people, but the cost of it went up and nobody in the room knows which line it landed on. The spreadsheet shows everything AI consumes and hides one thing only: what it started charging the people left over to check what it does.
That is the edge of FinOps in 2026. It sees what AI spends. It does not see what AI demands. And what it demands from your humans to close a decision charges far more than it ever consumed in tokens.
FinOps Has Two Waves. The Third Has No Name Yet.
The story of FinOps starts in an expensive misunderstanding. Engineering provisioned cloud capacity as easily as pressing a button, finance got the bill at month-end with no idea what had been bought, and the two sides eyed each other across the table. The FinOps Foundation, in 2019, gave that friction a shared vocabulary: visibility, optimization, operation. Whoever spends, let them see; whoever pays, let them understand. Five years on, any company with a serious cloud bill has someone whose job is to mind that number, and nobody finds it strange.
The second wave rode in on the flood of generative models. Same method, now pointed at inference, tokens, dedicated GPU, fine-tuned models, vector databases. CloudZero, Vantage, and Datadog shipped their AI spend visibility modules; OpenAI and Anthropic opened granular usage dashboards. In the more mature mid-market, that line already reaches the financial QBR without needing an explanation.
Both waves answer the same question: how much AI consumes while it runs as technology. Neither touches the other question: how much it costs while it operates as part of the workforce. And it is in that second one, the one nobody has named yet, that most of the aggregate bill lives.
| FinOps Reading | What It Covers | Unit | Year It Weighed on the Budget |
|---|---|---|---|
| 1. Cloud spend (traditional) | Compute, storage, network, product SaaS | Cost per workload per month | 2018 to 2020 |
| 2. AI spend (AI infrastructure) | Tokens, GPU, inference, fine-tuned models | Cost per call per model | 2024 to 2025 |
| 3. Human-agent coordination | Payroll in ratification, calibration, AI alignment meetings, handoff remediation | Cost per edge per decision | 2027 onward (projected) |
The third reading fits the same three pillars of the original Foundation, inventing nothing. Visibility first: map where human-agent coordination happens. Optimization next: find which edges grow faster than revenue. Operation last: a quarterly cadence, with payroll and cloud spend on the same radar. The method is the one you already know. Only the unit of measure changes.
The Real AI Cost Structure of a Mid-Sized Company
Executive instinct runs straight to the hosting bill: AI is expensive because inference is expensive, GPU is expensive, the model is expensive. Model the whole operation, not just the machine's invoice, and the proportion flips. Take a 500-person SaaS where most of the team already has an agent in the flow, apply a loaded senior hour to the human edges and average inference to the agentic ones, and the makeup of the total cost of the hybrid operation lands roughly like this.
| Category | % of Total Cost | Estimated Monthly Cost | Does Current FinOps Cover It? |
|---|---|---|---|
| AI infrastructure (tokens, GPU, model) | 8% to 12% | $70,000 to $110,000 | Yes, in AI FinOps tooling |
| Adjacent traditional cloud (storage, network) | 4% to 7% | $30,000 to $60,000 | Yes, in traditional FinOps tooling |
| AI-induced human coordination (H2H) | 45% to 55% | $400,000 to $530,000 | No, it sits under the payroll line |
| Senior ratification of what AI produced (A2H) | 18% to 26% | $160,000 to $220,000 | No, it sits under the payroll line |
| Prompt calibration (H2A) | 10% to 16% | $90,000 to $140,000 | No, it sits under the payroll line |
| Handoff remediation between agents (A2A) | 3% to 7% | $25,000 to $60,000 | Partial, it blends into senior eng payroll |
Look at the distortion head-on. AI infrastructure, the one thing today's FinOps sees, is the smallest slice. The four human-agent coordination edges add up to most of the cost, and they are diluted into payroll with no category to name them. Each one swells from a different trigger, which makes treating them as one block useless. The four edges, human with human, machine with machine, human with machine and machine with human, each one in currency are the vocabulary missing from the classic FinOps playbook.
Why Today's AI FinOps Does Not Close the Coordination Bill
AI FinOps tooling aims inside the call. How many tokens, which model, which team asked, which workload at which hour. It does that well, joins finance to engineering, and tells you precisely how much each product is costing in inference. Everything it sees sits on one side of the border: the machine's side.
On the other side, where it cannot reach, is where the expensive part lives. The analyst who hammered the prompt six times until the output was usable. The head who reviewed it, did not like it, and sent it back. The meeting born only to settle how the whole team is using the tool. None of that is an API call; it is people around the machine, and today's instrument has no way to grab it.
The effect is concrete. The report lands with a round inference line, say $90,000 for the month. You think it is high, sit down with the vendor, negotiate, switch models, cut 18%. Good execution, clean number. And completely irrelevant next to the hundreds of thousands the coordination around those same agents is burning every month in senior payroll, with no label, no line, nobody to add it up.
The FinOps Playbook Adapted for Coordination
The method does not change. The three pillars of the original Foundation, visibility, optimization, and operation, stay whole. What changes is the unit and the source of the data. In place of the call and the cloud workload, the unit becomes the decision that crosses the company and the edge it travels. In place of the vendor invoice, the source becomes the cross-reference between the senior calendar, what AI returned, and the loaded hour of whoever operates it.
| Pillar | Traditional FinOps | Coordination FinOps | Practical Difference |
|---|---|---|---|
| Visibility | Cost dashboard per workload | Edge inventory per decision | Draw the decision graph (human and machine) across 3 to 5 real cases |
| Optimization | Reserve capacity, right-size what is overprovisioned | Cut the edge that leaks most, split out the AI agenda | Drop the AI alignment meeting that could have been async, shorten the calibration cycle |
| Operation | Tag cost per team, alert when it passes the cap | Cost per decision, quarterly cadence next to payroll | A new line in the QBR, beside cloud spend and consolidated payroll |
The parallel holds for a practical reason: anyone who has already lived through the maturing of traditional FinOps needs to learn no new method. The mental toolkit is the same, the discipline is the same. What is missing is one new category inside a system you have run for years.
The Unit the Category Calls For: Cost per Decision
Traditional FinOps settled on cost per workload per month. AI FinOps stopped at cost per call per model. Human-agent coordination calls for a different ruler: cost per decision that crosses the company.
A decision that crosses the company is any cycle that starts with a question from leadership and ends in a ratified action. A big renewal. A pricing adjustment for one customer tier. A budget approval. In each, the decision passes through a few humans and a few agents, in a certain order, and the cost is the sum of what gets spent on every edge it travels.
The ruler is practical for three reasons. It fits the way the board talks: a decision is the word you and the rest of leadership already use to report results. It compares across teams: a pricing decision in sales has a shape much like a pricing decision in product. And it compares across companies. Model a typical decision like that in a mid-sized company in 2026 and it falls in a band of a few tens of thousands of dollars in loaded payroll. That is the calibration point, not a promise: if your decision costs well above it, there is structural fat in the path; if it costs well below, you are probably leaving context out of the ratification, and that bill comes back later, more expensive.
What Changes in the CFO's Room
The board that asks for a granular explanation of why margin did not keep up with the gains everyone swears they get from AI will want the answer in currency, not in adjectives. So you have three ways out. Say you have not measured it yet, and spend credibility. Guess from the inference report, and take the correction three months later when the aggregate bill shows up worse. Or put the new category on the table: human-agent coordination in dollars, set apart from the two waves that already exist.
This is the category where the AI savings are leaking into meetings with some regularity in 2026. Without it, any story about AI ROI stays half-told. With it, the CFO gets the vocabulary that was missing to explain the margin hole to the board without promising a miracle in three quarters.
Three Moves to Start Before Your Next QBR
No software needed. No consultant needed. Three moves deliver a defensible number in a few weeks, with the spreadsheet you already have.
- Inventory three recent decisions that crossed the company. Take three decisions that mattered in the last few months and ran through AI on the way. They do not have to be the three most prestigious, just representative. Draw the path of each one: how many humans, how many agents, in what order, how many round trips on each edge. It fits on one page. Across three decisions, your company's pattern already shows.
- Attach a loaded cost to each edge. The company's loaded senior hour on each human edge. Average inference on each agentic edge. An estimate of human remediation every time one agent hands the baton to another. The rough 15% to 25% margin serves for order of magnitude. Compare the decision total against the calibration band of a few tens of thousands of dollars for a typical decision.
- Take the category to the next QBR as a new line. No forecasting growth, no proposing cuts on the debut. Just show the category exists, what the unit is, and what it costs in order of magnitude. It is the same move the CFO used to present cloud spend back around 2017. First the company sees; acting comes later.
The Theory Anchors in 1937, Not in the FinOps Foundation
The FinOps Foundation gave the method for charging visibility on technology. But the idea that coordination is measured edge by edge is far older than the cloud. In 1937, Coase asked why firms exist, and answered: because coordinating inside the walls comes out cheaper than coordinating through the market, in certain cases. Williamson refined it in 1985, showing that this math depends on the transaction cost of each kind of edge. Measuring edge by edge comes from that economic tradition, not from a financial discipline born yesterday.
What 2026 brought that is new is a fourth structure in the graph of the firm: the AI agent operating inside without being a person, moving the cost of every human edge next to it. Coase and Williamson applied to 2026 pulls the whole foundation together. Coordination FinOps is just the translation of that old theory into today's CFO spreadsheet.
Frequently Asked Questions
What is Coordination FinOps?
It is the third reading of the FinOps discipline brought inside the operation. The first measures cloud spend: compute, storage, network. The second arrived in 2024 and measures inference: tokens, GPU, cost per call. The third measures what nobody put on a line: what AI charges your humans to close a decision. A senior reviewing what the model returned, a prompt that comes back five times before it is right, the meeting booked just to align who is using what, the remediation when two agents do not understand each other. That third reading has no budget category, no dashboard, no consolidated vendor. It has the biggest piece of the bill.
What is the difference between AI FinOps and Coordination FinOps?
AI FinOps counts what the machine consumes while it runs as technology: tokens, GPU, inference, cost per model, cost per call. A known category, with tooling ready. Coordination FinOps counts what the machine charges your humans while it operates as part of the workforce: the senior payroll spent ratifying, calibrating, and aligning. Model a 500-person SaaS at average adoption and AI infrastructure fits in a small slice of the total cost of the hybrid operation; the rest lives in the coordination edges, split across payroll with no category of its own. Without that reading, any AI ROI is a guess.
Why should a CFO care about this?
Three dry reasons. One: operating margin did not keep up with the gain everyone swears they get from AI, and the board will want an explanation in currency at the next QBR. Two: the bill costs a lot all year without showing up in the P&L, hidden between the payroll line and the cloud spend line, and what has no line cannot be governed. Three: deciding where to put AI capital without knowing which edge leaks most is betting in the dark. The CFO already governs cloud spend because one day it turned material and it hurt; human-agent coordination is on the same path, about five years behind.
How do you measure Coordination FinOps without new tooling?
Three moves cover the first month, and none of them ask for software. Take three recent decisions that mattered (a big renewal, a pricing change, a budget approval) and draw the path of each: how many humans, how many agents, in what order, how many round trips. Attach a cost to each edge, loaded senior payroll on the human ones, average inference on the agentic ones, with the rough 15% to 25% margin that serves for order of magnitude. Set it next to cloud spend and payroll on the same spreadsheet. In thirty days the category goes from anecdote to number.
Will human-agent coordination tend to earn a formal FinOps category?
The path of the two earlier waves suggests it will. Cloud spend took from 2015 to 2020 to earn a dedicated budget line; inference cut the path short in 2024 and 2025, when the invoice passed seven figures a year and became impossible to ignore. Human-agent coordination sits today where the cloud sat around 2017: a small team grasped the problem, the first vendors are starting to appear, the board is not asking yet. At the speed of adoption, the recognition tends to arrive faster than the cloud's, not slower.
The Bottom Line
The question that matters stopped being how much your company spends on AI. Now it is another one: how much it spends to coordinate human and machine, and whether that bill grows faster than revenue. Cloud spend took five years to earn a budget category. Inference cut it to two. Human-agent coordination, at the speed of adoption and the size of the leak, tends toward a similar horizon, and probably a shorter one.
In 2026 it is still the invisible vector of AI governance. At the pace of the two earlier waves, it reaches the QBR before the board asks and shows up in a serious vendor RFP soon after. Whoever takes the category first arrives at the capital conversation years ahead, talking in currency while the rest still talk in perception.
The instrument is still rough. The category is already measurable by hand. Deciding to measure now, on your own time, or to wait for the board to ask, is the kind of choice that separates who governs from who performs over the next eighteen months. The minimum defensible dashboard has five metrics that measure economic governance and five anti-metrics that only make noise.