AWS Just Shipped an AI Agent That Hunts Your Cloud Waste — and Smartly Refuses to Touch Anything

If you run anything on AWS, your cloud bill is probably the line item nobody fully understands and everybody is afraid to touch. This week AWS aimed an AI agent directly at that problem.

On June 15, AWS featured its new FinOps Agent in the weekly roundup, after putting it into public preview a few days earlier (AWS). It is a purpose-built agent for FinOps practitioners and engineering teams that answers cost questions in plain language, surfaces savings, investigates anomalies, and runs recurring cost workflows on a schedule (AWS Cloud Financial Management).

Here is what it actually does. You can ask it why last month’s bill jumped, and it will dig in. It pulls rightsizing, idle-resource, and Savings Plans recommendations straight from AWS Cost Optimization Hub and AWS Compute Optimizer, and each recommendation comes with the affected resource, the current and recommended configuration, the estimated monthly savings, the implementation effort, and guidance on how to do it (AWS). When it spots a cost anomaly, it can investigate the root cause on its own and post the findings to a Slack channel, or open a Jira ticket so a human picks it up (AWS). During the preview it carries no additional charge, though it is not yet available in GovCloud (US) or the AWS China regions (AWS Cloud Financial Management).

Now the part I want to flag in bold, because it is the most important design decision here: the agent is read-only. It investigates, it reports, it files a ticket, it sends a Slack message. It does not stop, resize, delete, or remediate anything (Cloud Cost Clinic).

I think that is exactly right, and I want to explain why.

What this means for you if you run enterprise IT

You already know the FinOps drill. Somebody exports a Cost Explorer dump, somebody else builds a spreadsheet, a third person nags engineering about the orphaned volumes nobody will delete, and three weeks later you do it again. It is real work, it is recurring, and it is exactly the kind of toil an agent should eat.

A scheduled agent that runs the cost review for you, catches the anomaly the day it happens instead of at month-end close, and drops a tidy summary into the team’s Slack — that compresses a process that used to take a FinOps analyst days. AWS pitches this as "continuous" cost governance instead of a monthly fire drill (Devoteam), and that framing is correct. The value is not the AI being clever. The value is that the boring review now happens every day without a human remembering to do it.

But read the recommendations before you act on them. Independent reviewers who tested the preview point out the obvious gap: the agent surfaces a "rightsizing" recommendation, but it does not know that the box it wants to shrink is your month-end batch server that only spikes for six hours (Cloud Cost Clinic). Context lives in your head, not in Cost Optimization Hub. So the agent is a phenomenal research assistant and a terrible autopilot, which is precisely why AWS did not give it an autopilot.

What this means for you if you run a small business

You probably are not running a FinOps team. But if you have anything meaningful on AWS — a SaaS product, a few EC2 instances, an RDS database — this is the kind of thing that quietly saves you a few hundred dollars a month you did not know you were burning. Idle resources and the wrong instance size are the two most common ways small operators overpay, and this agent is built to find both. It costs nothing during the preview, so the only investment is the hour it takes to set it up.

My take

I have not personally run the FinOps Agent — I do not run a large AWS footprint, and I am not going to pretend otherwise. So treat this as analysis, not a hands-on review.

But the design philosophy lines up exactly with what I have learned using Claude Cowork and Microsoft Copilot Cowork every day. The agents I trust most are the ones that draw a hard line at the dangerous action and hand it back to me. Claude Cowork, on my own machine, refuses to log into things for me and asks before anything sensitive. I have come to see that visible friction as a feature, not a bug. AWS made the same call here: let the agent do all the tedious investigation, and stop cold at the moment where a wrong move would cost you real money or real uptime.

That is the template for deploying agents in 2026. Automate the analysis. Keep a human on the trigger. The companies that hand agents the keys to "go fix it" unattended are the ones writing the cautionary blog posts six months later.

If you are on AWS, turn it on this week. Let it watch your bill. Just do not let it touch the dial yet.

News commentary by Brad Rowland — IT Infrastructure and Operations leader, automation builder, and AI implementer. Sources are linked inline.