Another Month, Another Amazon Web Services Frontier Agent Arrives

Nobody has ever opened AWS Cost Explorer and felt the elation of joy they are about to be presented with. Let’s be honest it’s the only dashboard I know where the loading spinner has time to make you nervous before the page has even rendered. You scroll, you filter, you squint, you filter some more and then you find the line item that’s doubled overnight, and then the real work starts… finding out which human did it, and whether they’re going to admit it, this feels like a scene from the Dark Knight with Cristian Bale bellowing in his deepest voice…. “Where did the money go? … tell me where it is?”

I’ve spent a fair amount of time around clients and their environments on the same hunt. Not because the data’s missing, AWS gives you a hosepipe of cost data via Cost Explorer, Cost Anomaly Detection, Cost Optimization Hub, Compute Optimizer. It’s all there, lined up like evidence on a board. What’s missing is a good detective. Someone still has to read the alert, chase the engineer, and turn “RDS went up 40%” into a sentence a CFO can nod along to in a meeting that started five minutes ago.

So, AWS built a detective. Or that’s the pitch behind the FinOps Agent, announced at FinOps X in June. It is an agent which answers cloud cost questions in plain English, investigates anomalies without being asked twice, and produces the report before anyone’s had to chase anyone for it.

I’ve had a proper poke around the preview. Here’s the technical version of what it does, where the seams show, and whether it earns a place next to the cost tooling you’re probably already running.

Strip the keynote announcement language and it’s a Bedrock-built agent which integrates with your existing hosepipe of AWS cost data, such as Cost Explorer for the numbers, Cost Anomaly Detection for alerts, Cost Optimization Hub and Compute Optimizer for recommendations, and CloudTrail for working out who did what and when.

It brings clarity to all these FinOps data sources and surfaces the information an engineer needs in the way a traditional FinOps practitioner (or detective) would. It’s your new interface to AWS cost management and FinOps.

Setup is quick. Switch to us-east-1 (that’s the only region the control plane runs in for now, though the cost data it can see covers every region you operate in bar GovCloud and AWS’s China regions), create the agent, click through a one-click Identity Access Management role setup, optionally wire up Jira and Slack, done. No professional services engagement required, which I say as someone whose job occasionally involves professional services engagements.

The secret sauce is the context files the agent reads you upload your account-to-owner mapping, your team definitions, your tagging conventions, and the agent starts resolving “what’s Team “X” spending” into the right accounts without anyone typing an account ID. When you then work with the agent, it has codified your AWS and business structure, and understands the context it is working within. But it really needs the investment upfront to get the best from it.

This is what saves you time and money, rather than being another cost visualisation tool.

Every org of a decent size has Cost Anomaly Detection switched on, and every org of a decent size also has a Slack channel where those alerts go to die quietly, maybe even a ITSM platform to create a statistic. Someone glances at it eventually, usually after the anomaly’s already sat on the bottom line costing you for three days, before the calls of “is anyone looking into this.”

This is where the Agents automation comes into play. You can setup automation so when an anomaly fires, it goes through the relevant CloudTrail activity, lines the timing up against the cost change, and comes back with a likely root cause and the account or team responsible. You can set a dollar threshold, so it doesn’t bother you over a £4 blip, and it routes the finding straight to Jira or Slack, addressed to the right team instead of a shared inbox nobody admits to owning. This automation example is easily 30 mins of a FinOps practitioners time saved.

Somewhere in every organisation there’s a recurring report that exists purely out of obligation. The AWS FinOps Agent is there to work as your personal FinOps analyst so you no longer must create the report you dread.

Schedule it daily, weekly, monthly, or however you report on cloud costs and governance, and it produces a presentation-ready in HTML, PDF or PPT. Not a CSV someone then has to dress up before a board meeting. An actual slide deck.

Pair that with the recommendations slide and it gets properly useful. Rightsizing, idle resources, Savings Plans, all pulled straight from Cost Optimization Hub and Compute Optimizer, and instead of sitting in a console nobody opens voluntarily, the agent can summarise them into a Jira ticket and assign it to the team that owns the resource. The optimisation opportunity turns up where the engineer already works, with the context attached, rather than as a link, they’ll bookmark and never click.

The most underrated use case, mostly because it sounds like the boring one.

An engineer doesn’t need to learn Cost Explorer’s filtering logic, which is like a Chinese puzzle, they just type “why did my costs go up last month” into the web app and get back the actual change, the services behind it, and the usage pattern that caused it. No report to build first. No asking the FinOps team how to use the tool. No new PowerBI dashboard.

Once the context files are loaded, it gets conversational. “What’s Team ‘X’ spend this quarter” resolves to the right accounts on its own, and a follow-up question holds the thread the way a colleague who already knows your environment would, rather than making you start over from a blank query box every time.

The real shift is who gets to ask. Cost questions stop routing exclusively through whoever on the FinOps team happens to know where to look, and your specialists get their time back for the work that needs a human brain attached to it.

The question I always ask before letting any AI agent near production billing data is “what’s the worst it can do?”

Here, not much. It’s read-only by default through customer-managed IAM roles, and the only write actions available are creating a Jira ticket or posting a Slack message, both of which you have to explicitly enable. It can’t touch tags, infrastructure, or billing configuration. Every call it makes shows up in CloudTrail, so there’s a full record sitting alongside the investigation findings themselves.

AWS has also been refreshingly upfront that this is deliberately human in the loop.  The agent surfaces findings and recommendations, you then decide what happens. Given the trust most finance teams place in anything that touches the bill, that’s the correct call and follows the existing FinOps processes most organisations have in place.

It’s a preview, and it acts like one, so a few honesty flags before you put this in front of a client.

Re-authentication has been flaky for some people, with sessions dropping unexpectedly. The Slack integration isn’t fully bedded in yet. There’s no support yet for connecting third-party HTTPS endpoints beyond Jira and Slack, and it’s not listed in AWS Marketplace, which matters if your procurement team likes everything tracked that way. Nothing here is a reason to wait. It’s the normal texture of a service weeks into public preview, and AWS tends to close these gaps fast.

I have experienced a few eyebrow raising moments while watching it do its thing. When it is building queries, you sometimes see syntax errors which I find peculiar, clearly it is no Claude Code when it comes to this. I have also found that it can get give the wrong context in the data it pulls back, for example, I asked it to look back over the last 12 months of costs and look at where my biggest costs had increased, and it flagged that in the first month my KMS costs had escalated 400% in one month, but when I went to investigate there was no huge number of KMS keys created in that month. With a bit more querying it turned out to have made a mistake with the dates and the jump in KMS keys which had happened the month previously, so this was most likely related to the billing cycle when the keys were generated. It did eventually get to the right answer after a few more prompts to tease the information out.

Worth saying plainly, because the keynote framing makes it easy to assume otherwise.  This is not a replacement for your enterprise-grade FinOps tool.  Pitching it as such would be an expensive mistake to walk back from later.

There’s no allocation engine here, no chargeback model, no cross-cloud view, and it’s AWS only. It leans entirely on whatever tagging and account structure you’ve already built, good or bad. If you need proper cost allocation and showback across business units, or you’re managing a multi-cloud estate, that foundation still has to come from somewhere else. It still relies on the native AWS cost optimisation recommendation engine, whereas leading third-party tools go deeper in surfacing optimisation options.

What the agent does well is sit on top of the native-AWS FinOps tooling foundation and make it conversational and fast. It’s the interface layer, not the platform underneath it. Conflating the two is how vendors end up overselling something and FinOps practitioners end up cleaning up the expectation gap afterwards.

Yes absolutely, with one condition don’t aim it at your entire estate on day one.

Pick the one recurring piece of cost admin that nobody enjoys, the QBR deck that eats an afternoon every month, or the anomaly alerts that sit unread until someone gets shouted at and let the agent just operate that for a month. If it earns its keep there, it’ll earn a wider rollout. If it doesn’t, you’ve lost very little finding out, because right now it costs nothing to try.