What AI agents should and shouldn't do in FinOps

In brief

Agents suit FinOps work that is read-only or easy to undo. Commitments, policy and chargeback disputes need a person. Where the line sits, with guardrails.

Steve Elliott

Steve Elliott · · 7 min read


On this page
  1. Why is AI for FinOps a priority now?
  2. What do the cloud providers' agents actually do?
  3. Which FinOps tasks suit agents?
  4. Which tasks need a person?
  5. What guardrails should a FinOps agent work within?
  6. Sources

In our view, AI agents should do the FinOps work that is read-only or cheap to undo: triaging anomalies, assembling recommendations, proposing tag fixes and drafting reports. Commitment purchases, policy and chargeback disputes should stay with a person, because they lock in money for years, set rules other people must follow, or settle disagreements between teams.

The test we propose is simple: what does a wrong answer cost, and can you take it back? This post applies that test to the agents AWS, Microsoft and Google documented in 2026, and sets out the guardrails we think every FinOps agent needs. It is about running a FinOps practice. Agents that keep a TBM cost model current are covered in our post on living IT cost models.

Why is AI for FinOps a priority now?

The FinOps Foundation's State of FinOps 2026, released February 19, 2026, surveyed 1,192 respondents. FinOps for AI, managing the cost of AI, is the top forward-looking priority. AI for FinOps, using AI to do the work, also ranks among respondents' top future priorities, and they rate its importance above the midpoint on average.

The report lists early use cases: anomaly detection and faster alerting, automated rightsizing recommendations, natural-language queries of cost data, automated discount procurement and resource tagging. It reads early signals as AI being "a capability amplifier, not a substitute for FinOps expertise."

Practitioners are cautious about letting agents act. A March 2026 FinOps Foundation article on agentic use cases, drawn from the author's conversations with practitioners, says most organizations are not yet comfortable letting agents "make production changes or delete resources without human approval." The Day 2 keynote at FinOps X 2026, titled "From Alerts to Agents", included a crawl, walk, run model for moving from automated alerts toward agentic cost management.

What do the cloud providers' agents actually do?

In our reading of the documentation as of November 2026: mostly read cost data, explain it and prepare work for a person. The vendors' announcements use stronger language, noted below. None of the provider pages we read describes an agent that buys commitments.

AWS FinOps AgentAzure Copilot Optimization AgentGoogle Cloud Billing (Explainability Agent, AI Cost Summary Agent, Gemini Cloud Assist)
StatusPreview since June 9, 2026PreviewExplainability Agent in console per April 2026 announcement; no launch stage stated for any feature
What it doesAnswers cost questions, investigates anomalies, summarizes recommendations, produces scheduled reportsRecommends changes to virtual machines and scale sets, explains why, compares alternativesExplains AI cost drivers; Gemini drafts emails on wasted usage and recommended optimizations, and builds reports
Actions outside itselfCreates Jira issues and comments, posts to SlackGenerates Azure CLI or PowerShell scripts that you can deployReport filters; email drafts
ApprovalRequired for Jira actions started from chat or on demand; pre-authorized in scheduled automations; none for SlackNot documented on the agent pageNot described

The AWS documentation is the most explicit. Read-only actions, such as querying cost data and retrieving recommendations, run without approval. The agent summarizes recommendations from Cost Optimization Hub and Compute Optimizer but does not generate its own optimization analysis. AWS puts responsibility for decisions and actions taken on the agent's output on the customer, and recommends human review.

Microsoft's February 2026 announcement says the Optimization Agent identifies and executes improvements across cost, performance and sustainability, and that every agent-initiated action honors existing policy, security and RBAC controls. The current Learn page describes recommendations and scripts you deploy. Google's April 2026 announcement says its FinOps Explainability Agent is "designed to operate autonomously" to investigate the drivers of AI-related cloud costs; it doesn't say the agent changes resources.

Which FinOps tasks suit agents?

Tasks where the agent's output is a draft, a ticket or an explanation, and a mistake costs someone's time rather than a contract.

  • Cost anomaly triage. The AWS agent correlates a cost anomaly with CloudTrail events and names a likely root cause and owner. Its own documentation tells you to treat that analysis as "a starting point for investigation rather than a definitive conclusion." We think that is the right frame. In our view a wrong diagnosis costs an analyst some time, while a correct one shortens the follow-up.
  • Recommendations and remediation tickets. Agents can gather rightsizing and idle-resource findings, find the owner and open a ticket. The FinOps Foundation article describes one team cutting initial investigation from 15 minutes per ticket to essentially zero.
  • Tagging fixes. Practitioners are exploring agents that work out what unlabeled resources are from when they run and what they connect to. We think the agent should propose cost allocation tags and the owner should confirm them, because a wrong tag moves cost into another team's showback.
  • Report drafting and cost questions. Azure Copilot answers questions such as forecasting cost by month, and the AWS agent produces scheduled reports in HTML, PDF or PPT. AWS's documentation asks you to check that charts match the underlying data, which is a fair rule for any agent-drafted report.

Which tasks need a person?

Tasks that commit money you can't easily recover, set rules for others, or decide between people.

TaskWhat an agent can doWhat a person should doWhy
Commitment purchasesModel coverage, draft the purchase caseApprove the strategy and every purchase outside a pre-approved limitCommitments run one or three years, and refunds are limited
PolicyDraft policy text, report complianceSet the policy and approve exceptionsPolicy needs authority behind it
Chargeback disputesAssemble the data behind a chargeDecide the outcome with financeIt moves cost between budgets and the ledger
Deleting production resourcesFind candidates, open ticketsApprove the deletionSome deletions can't be undone

Commitments. An AWS Savings Plan is a one- or three-year commitment. It can be returned only if its hourly commitment is $100 or less, it was bought in the last seven days and in the same calendar month, and the account hasn't hit its return limit. The FinOps Framework's Rate Optimization capability notes that the organization still pays when no matching resource is running.

On Azure, as of November 2026, refunds on canceled reservations are capped at $50,000 of commitments per billing scope in a rolling 12-month window. From February 1, 2027, new reservations for services covered by savings plans can't be exchanged; reservations bought before then keep one final exchange. Microsoft doesn't charge an early termination fee today but says it might in future.

Take an illustrative three-year compute savings plan at $150 an hour. That is $1.31 million a year and $3.94 million over the term, and at that size it can't be returned at all. A plan at $40 an hour commits $1.05 million over three years and can be returned only within seven days, in the same calendar month.

The State of FinOps lists automated discount procurement as an early use case, and the Rate Optimization capability describes mature practices running frequent purchase cycles triggered by automated reporting. It also lists modeling pre-approvals among finance's activities. In our view the line is between automation inside an envelope a person approved in advance and an agent deciding the commitment strategy. The first is sound practice. The second puts a multi-year obligation on a probabilistic system.

Policy. The FinOps Foundation's Governance, Policy & Risk capability says leadership sets governing principles that FinOps policies turn into operational rules. Its examples include a defined exception pathway, such as a documented business justification, reviewed and approved, before an idle resource escapes decommissioning. In our view agents can draft policy text and enforce a policy once it exists, but they shouldn't set the rule they then enforce.

Chargeback disputes. The Invoicing & Chargeback capability says FinOps and finance decide together whether costs are held centrally or allocated, and calls for a process to correct chargeback reporting after the period closes. It doesn't describe a dispute process. In our view a dispute is a decision about whose budget absorbs a cost: an agent can bring the evidence, and finance and the budget owners decide. Our post on moving from showback to chargeback covers the policy side.

What guardrails should a FinOps agent work within?

AWS sets out four tiers of automation, from read-only insights through simple mutation and rule-based automation to autonomy with boundaries. Its advice: "Start with a read-only insight workflow today." We agree, and we think four guardrails should apply at every tier.

  1. Approvals at the point that matters. Because AWS pre-authorizes scheduled automations at setup, the real approval is often the setup. We think it deserves the same review as a change request.
  2. Reversibility. AWS's trust post says automated actions should be easy to reverse, citing a deleted EBS volume restored from a snapshot taken before the automation ran. If an action can't be undone, a person approves it.
  3. An audit trail. AWS says CloudTrail captures all FinOps Agent activity, and the agent's reasoning steps, tool calls and task history can be reviewed. Microsoft says agent actions are "always reviewable, traceable, and auditable." In our view the log should show the data the agent used, what it recommended and who approved it.
  4. Spend and scope limits. AWS describes boundaries set by account, region or maximum financial impact. Deterministic caps help too: Google made spend caps on Cloud Budgets generally available on October 30, 2026. For eligible services they halt pay-as-you-go and CUD-covered requests once a budget is hit, without deleting resources, though Provisioned Throughput keeps running. Budget the agent's own cost as well; the FinOps Foundation article suggests tracking a token budget per project.

The questions these guardrails answer are who decided, with what authority and from which data. That is where we think an operating model earns its keep: it connects each automated action to the owner, budget and policy it acts on.

Sources

Frequently asked questions

Should AI agents buy cloud commitments?

Not on their own judgment, in our view. Agents can model coverage and draft the case, and rule-based automation can buy within limits a person approved in advance. The strategy and anything outside those limits should stay with a person: AWS Savings Plans run one or three years, and only plans of $100 an hour or less can be returned, within seven days and in the same calendar month.

What can the AWS FinOps Agent do?

As of November 2026 it is in preview. It answers cost questions, investigates anomalies by correlating cost changes with CloudTrail events, summarizes recommendations from Cost Optimization Hub and Compute Optimizer and produces scheduled reports. Its only actions outside itself are creating Jira issues and comments and posting to Slack. AWS says it doesn't modify raw cost data.

Which FinOps tasks are safest to hand to an agent first?

Read-only work. Anomaly investigation, cost questions in plain language, recommendation summaries and report drafts carry little risk because a person still acts on the output. AWS recommends starting with a read-only insight workflow. Tag suggestions come next, with the resource owner confirming each one, since a wrong tag moves cost into another team's showback.

What guardrails do FinOps agents need?

They need four, in our view: approvals for anything that changes resources or money, including the automation setup itself; reversibility, so irreversible actions always go to a person; an audit trail showing the data used, the recommendation and the approver; and spend and scope limits, such as caps by account or financial impact, plus a budget for the agent's own running cost.

Free playbooks

In-depth guides for product and operations leaders.

Browse the playbooks
Steve Elliott

Steve Elliott

Published November 22, 2026

More on FinOps & Cloud

FOCUS, commitments, shared cost.