AWS FinOps Agent finds the cost. Knowing what to fix, and fixing it, is the hard part

What the FinOps Agent does in preview

The agent runs in US East (N. Virginia) and reads from the cost services AWS already operates: Cost Explorer, Cost Anomaly Detection, Cost Optimization Hub, Compute Optimizer and CloudTrail. Set-up is one agent and one IAM role, with optional Jira and Slack connections.

Preview use is free up to a monthly usage limit, with standard charges for the underlying services. AWS named Workday, AVIV Group, Convera and Mitre 10 as early users.

• Anomaly investigation: when Cost Anomaly Detection fires, the agent correlates the cost change with CloudTrail events, names a likely root cause and a probable owner, and writes a summary.
• Cost questions: engineers ask why a service cost more last month and get an answer in terms of the contributing services and usage drivers. Uploaded context files, such as an account-to-team mapping, help it answer team-level questions.
• Recurring reports: scheduled HTML, PDF or PowerPoint reports.
• Optimisation summaries: it pulls rightsizing, idle-resource and Savings Plans recommendations from Cost Optimization Hub and Compute Optimizer and writes them into a Jira ticket for review.

Where its job ends

Read the capability list again and notice the verbs: investigate, answer, summarise, deliver. The agent's output is a Jira ticket or a Slack post. The changes it makes to your account are the EventBridge rules that trigger its own automations. Every recommendation it surfaces is one that Cost Optimization Hub or Compute Optimizer had already computed. Its value is turning that data into readable findings and routing them to the right team.

That is the assessment half of FinOps, and assessment has been getting steadily easier for years. Cost Explorer, Storage Lens, Compute Optimizer and now an agent that reads them all and writes the memo. The part that has stayed hard is what happens after the memo lands.

The skill that stays scarce: knowing what to fix, and how

Take the most common finding on a storage bill: a bucket of business files is paying the S3 Standard rate for data nobody has opened in a year. The recommendation writes itself. The correction does not.

None of these questions is answered by a cost report. Each needs someone who knows how S3 bills, how lifecycle rules behave, what the business keeps, and how to apply a change in a way that can be undone. That person is the bottleneck in most teams, and the ticket from the FinOps Agent sits in their queue behind everything else.

• Which buckets? Intelligent-Tiering gains nothing on objects under 128 KB, so a bucket of a million small files should be left alone, and the monitoring fee would make it slightly worse.
• Which tier? Infrequent Access and Archive Instant Access open files immediately. Archive Access and Deep Archive Access take hours to restore, and a sales team that opens a 2023 contract once a quarter will notice.
• Which versions? Expiring old versions is permanent. Some buckets have Object Lock or a legal reason to keep history, and those have to be found and left out before the rule goes on.
• Which existing rules? Putting a lifecycle configuration replaces the whole configuration on the bucket. Apply a new rule without copying the old ones and you have quietly removed someone else's retention policy.
• Is the data actually cold? The bill shows data transfer, which is an upper bound on reads. Without access logs or a CloudTrail data trail, "unread" is an assumption, and assumptions are what delete the wrong thing.

The third task: proving the saving

There is a step after the fix that is just as often skipped. A change is applied, everyone assumes it worked, and nobody compares the next bill with the one before. Intelligent-Tiering moves objects to Infrequent Access after 30 days unread and to Archive Instant Access after 90, so the full saving takes a quarter to show up and arrives in stages. Version expiry shows on the very next bill. Unless someone holds a baseline and checks each month, the saving is a forecast, and forecasts have a way of being quoted as results.

A cost tool that stops at assessment leaves you with three jobs: decide, apply, verify. The decision needs expertise, the apply needs care, and the verification needs patience and a spreadsheet. The FinOps Agent removes the first afternoon of work and leaves the other three.