To my fellow FinOps practitioners, engineers, and leaders attending this year’s FinOps X conference in San Diego, I hope you have an amazing week of learning, connecting, and enjoying time with the community.

It is clear AI will be one of the dominant themes of the conference. And with the Linux Foundation’s recent announcement of the Tokenomics Foundation, tokens will likely be part of almost every conversation.

That is a good thing. The Tokenomics Foundation’s focus on open standards, benchmarks, and best practices for AI infrastructure economics is an important step forward. The industry needs a better way to measure, compare, and reason about token-based consumption.

But measuring AI infrastructure is not enough.

If agents are going to do real work inside companies, the cost is not just tokens. The cost is everything required for the agent to reach the work, understand the work, perform the work, and prove the work was done correctly.

That is the part we cannot forget.

OpenAI made this direction clearer recently, though it may have been overshadowed by the news that it confidentially submitted a draft S-1 to the SEC. Separately, recent reporting from TechCrunch described OpenAI’s push to make ChatGPT less of a chatbox and more of a gateway into agents, coding tools, and partner apps.

Read that as a product story and it is interesting.

Read it as a FinOps story and it is a major signal.

The chat interface is becoming less important than the operating layer around it. The gateway matters. The tools matter. The context matters. The orchestration matters. The agent is only useful if it can do work somewhere.

That raises the question most AI cost models still avoid:

Where does the agent actually work?

Until recently, nearly every internal system and third-party tool was designed for human operators. Dashboards assume someone will interpret what matters. Workflows assume someone will click through the right sequence. Approvals assume a person understands the context behind the request. Documentation assumes human judgment, memory, and inference. Most SaaS tools are still built around screens, filters, exports, and decisions made by people looking at the interface.

A person can open a dashboard and interpret what matters. A person can search Confluence and decide which page is current. A person can read a Slack thread and infer which answer is authoritative. A person can click through a SaaS workflow, pause when something looks wrong, and ask the right person before continuing.

An agent does not work that way. An agent needs a callable interface. That may be an API. It may be a CLI. It may be an MCP server. It may be a structured workflow exposed through an internal tool. But it cannot rely on the same assumptions that make human-facing software usable.

A human can tolerate ambiguity. An agent turns ambiguity into retries, wrong tool calls, longer context, failed workflows, and more tokens.

This is why Model Context Protocol matters. MCP gives systems a standard way to expose tools and context to AI applications. OpenAI’s Agents SDK, for example, points in the same direction: models become more useful when they can use tools, access context, and leave traces of what happened.

Existing systems need to become tool surfaces that support both humans AND agents.

A tool that is good enough for a human may not be good enough for an agent.

Human-facing tools can be messy because humans compensate. We interpret unclear labels. We retry manually. We notice when a result looks wrong. We ask follow-up questions. We understand that a dashboard, workflow, or approval process may not mean exactly what it appears to mean.

Agents need something different.

They need clear schemas, predictable outputs, structured errors, permission boundaries, latency targets, audit trails, and observability. They need retries that do not create runaway loops. They need idempotency, so the same action does not accidentally happen twice. They need enough structure to know what happened, what failed, and what to do next.

That is the mirror world.

Same company. Same systems. Same workflows. But translated into a form machines can use.

And that mirror world has a cost.

When teams estimate AI cost, they usually start with tokens. That is understandable because tokens are visible. But an agentic workflow creates cost across a much wider surface area.

There is the model call. There is the context loaded into the prompt. There is the tool call. There is the MCP server, API, or CLI. There is the retrieval layer. There is the vector database. There is multi-modal. There is the orchestration system. There are retries. There are evals. There is observability. There is human review when confidence is low. There is engineering time to build and maintain the whole thing.

None of that is optional if the agent is doing meaningful work.

This is the part most AI business cases will miss. The first version of the spreadsheet will show token cost. The real operating model will create cost across cloud, SaaS, data platforms, internal tools, developer environments, security systems, observability platforms, and workflow engines.

One agent used by one person may look manageable. One agent used by every employee starts to look different. Multiple agents used across engineering, finance, support, sales, operations, and product becomes something else entirely.

At that point, AI cost is no longer a single line item. It becomes a usage pattern across the technology landscape.

Every agent interaction has a cost. Every tool call has a cost. Every retry has a cost. Every evaluation has a cost. Every supporting system has a cost.

If those costs are not connected to the workflow, product, team, or business outcome they support, organizations will be back in a familiar place: spend is growing, everyone can see pieces of it, and no one can explain the whole thing.

This is where FinOps has to enter the conversation. Not because FinOps has all the answers, but because FinOps already has the right problem statement: make technology spend explainable, make ownership clear, make usage measurable, and connect cost to decisions.

That language already exists. Increasingly, the data language exists too.

FOCUS, the FinOps Open Cost and Usage Specification, gives organizations a common way to normalize cost and usage data across technology providers. Without a common language, each system describes usage differently. One reports tokens. Another reports requests. Another reports compute time. Another reports seats. Another reports API calls. Another reports workflow runs or credits.

That fragmentation is manageable when usage is small. It becomes a serious problem when agents become a standard interface to work.

FOCUS will not solve every nuance of agentic AI by itself. But I believe it gives organizations a foundation: a way to normalize cost and usage, compare across categories, and start connecting AI activity to the same financial operating model already used for cloud, SaaS, and broader technology spend.

The agent era does not need a brand-new financial language from scratch. It needs the existing FinOps language to expand into the agent layer.

So when someone proposes an agent, do not only ask what model it uses. Ask what systems it needs to call. Ask whether those systems have APIs, CLIs, or MCP interfaces that agents can actually use. Ask what performance standard those interfaces need to meet. Ask how retries, failures, permissions, and approvals work. Ask how token cost, tool cost, infrastructure cost, and human review cost will be measured. Ask whether those costs can be mapped to a product, team, workflow, or business outcome.

If the answer is no, the organization does not know what the agent costs. It only knows what part of the agent costs.

Tokens are easy to count. Will the mirror world be ready to incorporate this unit? Because, if agents become the way work gets done (which is a resounding Yes), this mirror world is where a large share of AI economics will live.

And FinOps needs to count it.

Written with the help of AI. All the ideas expressed are mine and mine alone.

If you would like your company included in the Spotlight, contact the CloudXray AI Team

Company: CloudXray AI

Category: Managed Services & Consulting

What They Do: FinOps Consulting & Advisory Services; Owners & maintainers of the single largest FinOps company directory (finops.cloudxray.ai)

Why It Matters: Companies still need guidance on implementing FinOps and understanding the landscape of companies that exist