Agentic FinOps
AI agents that investigate cloud and AI spend and carry out FinOps work within limits you set.
Agentic FinOps is the use of AI agents to do FinOps work: they read cost data, investigate changes, and propose or make optimizations within limits you set. The FinOps Foundation files it under AI for FinOps, which applies AI to the practice itself, and describes agentic AI as proactive and iterative, able to use tools across your technology stack to reach an outcome. A March 2026 Foundation review of practitioner use cases finds four recurring themes: eliminating manual data work, autonomous waste discovery, enforcing policy earlier in development, and personalized outreach. Agents reach billing data through APIs or the Model Context Protocol, an open-source standard for connecting AI applications to external systems. An August 2026 Foundation article sets out levels from Level 0, Not an Agent, to Level 4, Acts within Limits, and says only the final level has authority to take or initiate changes. LevelFour runs an MCP server over your spend and savings, so Claude Code, Cursor or any MCP client can read costs and start a rollout.
Frequently asked questions
What is the difference between agentic FinOps and FinOps automation?
FinOps automation follows steps a person scripted, and agentic FinOps lets a model choose its next step. A nightly script that flags untagged resources does the same thing each run. An agent decides which data to read and what to try next, within limits. A FinOps Foundation article from August 2026 makes this the test for any product sold as an agent: was the sequence written down by a developer in advance, or can the model choose its next action? You can combine the two, with an agent that investigates and scripted automation that applies the change.
What is the difference between agentic FinOps and FinOps for AI?
Agentic FinOps uses AI to do FinOps work, and FinOps for AI manages what you spend on AI. An agent that investigates a cost spike is agentic FinOps. A budget on your OpenAI and Anthropic bills is FinOps for AI. The FinOps Foundation covers the first under AI for FinOps and treats the second as its own topic. Teams that run agents need both, because an agent's model calls add to the AI bill. The Foundation advises a spending cap and a step limit for that reason.
What can a FinOps agent do today?
A FinOps agent can answer cost questions in plain language, investigate a spike, find waste and draft changes. The FinOps Foundation's March 2026 article lists natural language dashboards and financial reconciliation, autonomous waste discovery, proactive guardrails in the CI/CD pipeline, contextual resource labeling and personalized outreach. One practitioner quoted in that Foundation article reports an action rate of 40% to almost 50% on messages sent through Slack. The same article says agents serve today as co-designers or companions, so expect a person to stay involved in production changes.
Is it safe to let an AI agent change cloud resources?
Yes, if a written policy and scoped credentials bound what the agent can do. The FinOps Foundation advises you to define limits through policy rather than prompts, and to begin with investigation, because a wrong answer at that stage costs little. It also says to route the first changes through your existing review process. For writes, use short-lived permissions scoped to one change, and log each call. The Foundation's March 2026 article reports that most organizations aren't yet comfortable with agents deleting resources without human approval.
What role does MCP play in agentic FinOps?
The Model Context Protocol gives an agent a standard way to reach cost data and actions. MCP is an open-source standard for connecting AI applications to external systems, supported by clients such as Claude, ChatGPT, Visual Studio Code and Cursor. A cost platform exposes its data as tools, and the agent calls them. AWS Labs publishes a Billing and Cost Management MCP server whose API calls run with the caller's AWS credentials. A FinOps Foundation paper lists MCP use cases such as automated cost analysis, anomaly detection and cost impact simulations.
How do you start with agentic FinOps?
Start with one read-only question your team answers by hand each week, such as why a service's spend rose. Connect an agent to billing data with read-only access, and compare its answers with your own for a few weeks. The FinOps Foundation suggests beginning with investigation, then having the agent propose rather than perform, with a spending cap and a step limit on the loop. Grant write access last, one action type at a time, with an audit log and a way to undo.