I called this one
Two days before AWS shipped it, I wrote that the direction of FinOps X was unmistakable. AWS and Google were both pointing toward autonomous cost management : agents that don't just surface recommendations but act on them. I said the practitioners who understood what was coming had time to position accordingly.
On June 9, AWS announced the public preview of the FinOps Agent. I'm not bringing this up to take a victory lap. The point of writing about a trend before the product ships is that you understand the trend, not that you guessed a date. What matters now is what the Agent actually does, and whether it lives up to what AWS demoed on stage.
I didn't get preview access myself, mainly because there was nothing to get. The preview is open right now to any AWS account, free of charge, no waitlist, no approval step. What I have is something almost as useful as hands-on access: several practitioners turned it on in the first days and published exactly what they found, rough edges included. That's the version of this story worth reading, not the press release.
What AWS actually shipped
The FinOps Agent is an agentic AI solution built to investigate cost anomalies down to root cause and answer cost questions for engineers in the tools they already use — Slack and Jira specifically, not a new console you have to learn.
It runs on three triggers. A recurring schedule you define. The moment a cost anomaly is detected. Or on demand, whenever an engineer asks it a question. Under the hood it draws on Cost Explorer, Cost Anomaly Detection, Cost Optimization Hub, and Compute Optimizer : the same data your central FinOps team already works from, which matters because it means the Agent isn't introducing a second source of truth that contradicts your existing dashboards.
A few practical details worth knowing before you evaluate it. The agent itself currently runs only out of US East (N. Virginia), though it can pull cost and usage data covering every AWS region except GovCloud and the China regions. It's free to use during the preview period, though the standard charges still apply to whatever AWS APIs it calls internally, Cost Explorer being the main one. And it's genuinely open right now, not gated behind a waitlist or an approval process — any AWS account holder can turn it on today.
The capabilities that matter most for day-to-day FinOps work: it surfaces rightsizing, idle resource, and Savings Plans recommendations directly from Cost Optimization Hub and Compute Optimizer. It can open Jira tickets on your behalf based on those recommendations. And when an anomaly fires, it investigates the root cause automatically and posts findings to Slack, rather than waiting for a human to start digging through CloudTrail.
Early customers named in the announcement include Workday, AVIV Group, Convera, and Mitre 10. AWS's framing is that this moves FinOps from reactive monthly reviews to scheduled, event-driven operations.
That's the pitch. Reasonable pitch. The question, as always with a preview, is what happens when it meets a real production environment with real account sprawl and real edge cases.
What practitioners found when they actually used it
This is where the story gets more useful than the announcement.
One practitioner deployed the Agent in the first days of preview specifically to answer one question: beyond the announcement, what does this actually change for a team doing FinOps day to day? The first observation was that getting started is remarkably quick — which matches AWS's pattern with most of their recent CFM tooling. Low setup friction has been a consistent strength across Cost Explorer's natural language features, the Bedrock attribution work, and now this.
But the same review listed a specific set of gaps that any practitioner evaluating this for production use needs to know before getting excited.
The Slack integration has real setup friction. One reviewer got everything configured, only to hit an error because the FinOps Agent app needed to be manually added as a member of the Slack channel before it would actually work, with no upfront explanation that this step was required. Another reviewer found they had to disable multi-session in the AWS Management Console before agent permissions would configure correctly, again with no warning ahead of time. Neither of these is a fundamental flaw. Both are the kind of friction that shows up consistently in week-one preview software and gets ironed out once enough people hit the same wall and file the same support ticket.
Integration with third-party HTTPS endpoints isn't available. If your organisation runs anything outside the AWS-native toolchain that you'd want the Agent to reach into, that's not on the table yet.
Re-authentication doesn't work reliably. The reviewer documented unexpected disconnections during testing. For a tool meant to run on a recurring schedule and respond instantly to anomalies, an auth layer that drops sessions unexpectedly is a real reliability concern, not a cosmetic one.
Context files only accept a restricted set of formats. You can't, for instance, hand the Agent an architecture diagram as an image to give it visual context about your environment. For a FinOps practitioner who wants the Agent to understand why a particular service exists and how it connects to others, that's a meaningful limitation right now.
None of this means the product is bad. It means it's a preview, behaving like a preview. The demo on stage at FinOps X showed the capability working cleanly because demos are built to show capability working cleanly. The actual deployed experience, six days after the demo, has the friction you'd expect from a feature that's weeks old.
What this means if you're evaluating it now
Here's my honest read, and where I'd push back gently on the more breathless coverage of this launch.
The autonomous direction is correct. I said that after Day 1 of FinOps X and nothing about the preview experience changes that view. Moving routine anomaly investigation and recommendation surfacing into an automated, always-on layer is the right direction for the discipline, and AWS building it directly into Slack and Jira rather than a separate console is the right design decision. Practitioners spend real hours every week doing the manual version of exactly what this Agent automates: noticing a spike, opening Cost Explorer, filtering by service and account, jumping to CloudTrail, tracing it to a cause. Compressing that into a scheduled or anomaly-triggered workflow is genuinely valuable.
What I'd push back on is treating this as production-ready today. If you're a FinOps lead deciding whether to roll this out to your engineering teams this month, the honest answer based on what's been reported is: not yet, and specifically not yet for the Slack-first workflow that is the entire point of the product. The re-authentication issue alone is enough reason to keep this in a sandboxed evaluation rather than a production rollout where engineers are relying on it during an active incident.
What I would do, starting now: get into the preview, run it against a non-critical account or a dev environment, and specifically stress-test the things that were flagged as broken. See if the auth issue reproduces in your environment. See how the anomaly investigation quality compares to what your team would have found manually. Build your own honest assessment rather than relying on the demo or, frankly, relying entirely on this article.
The gap between an AWS product announcement and a production-ready capability has historically run somewhere between two and six months for agentic features in this space. Amazon Q in Cost Explorer launched in April with real capability and real gaps, and most of the gaps closed within two to three months. I'd expect a similar curve here. Preview now, pilot in Q3, consider production rollout once the Slack integration and auth reliability are confirmed stable by people other than AWS's own marketing.
The bigger pattern this confirms
Stack this against what else has shipped since FinOps X closed. Google announced Automated Spend Caps the same week. IBM shipped a FOCUS AI Agent. Flexera announced autonomous rate optimisation through ProperOps+. AWS's own Agent is the fourth major "autonomous cost action" announcement from a different vendor within about ten days of each other.
That's not coincidence. That's an entire industry category forming in real time, with every major player racing to ship the same category of capability within the same month. When that happens, the early releases are reliably rough. Nobody ships a mature, fully-tested agentic product in a category that didn't exist a year ago, in a race against three competitors all announcing in the same window.
The practical implication for FinOps practitioners: this is the right moment to start learning how these tools work, what questions to ask of them, and how to design the governance layer that sits above autonomous action. It is not yet the moment to hand them the keys to production cost management without supervision. Those are two different timelines, and conflating them is how a genuinely good idea turns into an expensive lesson about why you test things in dev first.
Evaluating AWS FinOps Agent or any of the new autonomous cost tools for your environment? Get in touch — the pilot design and governance guardrails are usually the harder part, not the initial setup.