The Dawn of Agentic FinOps

Over the past few years, FinOps has matured into a discipline that most cloud-native organizations now recognize as essential. Cost visibility, allocation, forecasting, and accountability are no longer optional. They are part of how modern engineering and finance teams operate together.

At the same time, something else has been quietly changing.

AI workloads are no longer isolated experiments. They are becoming embedded into products, internal tools, and decision flows. In some cases, they are starting to operate semi-autonomously through agents that plan, execute, retry, and adapt without human intervention at every step.

This blog series exists at the intersection of those two shifts. It is not a declaration about what FinOps “should be” in 2026. It is a practical exploration of what changes when AI systems become part of the cost surface, and what I have learned while trying to manage that reality in practice.

Why this series

Traditional FinOps practices assume a certain rhythm. Usage happens first. Costs are reported later. Decisions follow.

AI disrupts that sequence.

With large language models and agentic systems, cost is incurred at the moment a decision is made. A longer prompt, an extra retry, a different model choice, or a poorly bounded loop can materially change spend in seconds, not weeks. These decisions often live deep inside application logic, not infrastructure configuration.

This does not make existing FinOps practices obsolete. But it does mean they are no longer sufficient on their own. What I am trying to understand, and share through this series, is how FinOps evolves when cost becomes a runtime concern rather than a reporting artifact.

From visibility to systems

One of the early lessons in this journey was that better dashboards were not the answer. AI costs can spike, decay, and shift ownership faster than traditional reporting cycles can capture. By the time a monthly report is reviewed, the system that caused the spend may already have changed behavior, or moved on to a different model entirely.

This pushed me toward a different framing. Instead of asking how to report AI costs more accurately, I started asking how systems themselves could be designed to make better cost decisions. That includes how requests are routed, how context is managed, how memory is persisted, and how guardrails are enforced when no human is in the loop.

This is where the idea of Agentic FinOps started to take shape.

Not as a replacement for FinOps teams, but as a way to embed financial intent directly into AI-driven systems.

A personal journey, not a finished doctrine

I did not start this work with a grand architecture in mind. It began with concrete problems: AI experiments that quietly became production workloads, costs that were hard to explain after the fact, and agents that behaved correctly from a functional perspective but poorly from an economic one.

Each project, prototype, and misstep added a bit more clarity. Some of that clarity is reflected in the tools, assessments, and code I am building along the way. Some of it shows up as questions I still do not have full answers to.

This series is an attempt to make that learning explicit. To document what seems to work, what breaks down, and what needs to be designed earlier than most teams expect.

What you will find in this series

The posts that follow explore FinOps for AI from several complementary angles. Some focus on how AI cost signals behave differently from traditional cloud workloads. Others look at why real-time intelligence matters more than post-hoc analysis. Several dive into architecture, including why MCP - Model Context Protocol and agent-based patterns change how governance and policy enforcement can work.

Later posts move into unit economics, ROI, and what it means to reason about value when every inference has a marginal cost. One of the posts looks at how an agentic FinOps architecture can be assembled in practice, and what trade-offs it introduces.

If you are responsible for cloud cost management, AI platforms, or both, my hope is that you will find something here that helps you think more clearly about where FinOps fits in an AI-driven world. Not as a static framework, but as a discipline that is learning to operate inside systems that now make decisions on their own.

This post introduce the AI for FinOps series:

1. Why traditional FinOps breaks with AI workloads: a story of costs, tokens, and fat tails

2. Why AI costs management require Real-Time Intelligence, not another dashboard

3. Why Model Context Protocol is the runtime for Agentic FinOps

4. Why MCP is the pivot for cost and policy governance (to be published)

5. A practical ROI framework for AI workloads (to be published)

6. Designing a FinOps agent architecture (to be published)

7. Memory and context management in FinOps agents (to be published)

8. Are you ready for Agentic FinOps? A practical maturity model (to be published)