AWS FinOps Agent vs Sherpa: what a cost-only agent can't see

AWS shipped a FinOps Agent into public preview last week — ask your cloud about its costs in plain language. I wasn’t worried. I was glad. When a hyperscaler ships into your category, the category stops needing a sales pitch.

So we pointed AWS’s FinOps Agent and Sherpa at the same AWS account and asked the same questions. Here’s where the two agents tie, where they differ, and the running cost AWS doesn’t put on the label.

Where the two agents agree

Start here, because it’s the honest foundation: where AWS has a recommendation, Sherpa matches it. We asked both to find optimization opportunities and the rightsizing numbers came back identical, to the dollar. When two independent agents read the same environment and land on the same figure, you can trust both are reading reality. No hand-waving. That’s the baseline.

The waste that has no “recommendation type”

Then we totaled it up. On our own test account, AWS’s agent came back with roughly $141/month in savings across four categories; Sherpa came back with roughly $180/month on the same account — and not from a cleverer prompt. AWS’s “find my savings” answer is built from its cost recommendation engines, so it returns the categories those engines model and nothing else. The rest is invisible to it: idle load balancers, NAT gateways doing almost nothing, IP addresses you’re paying for and not using, over-provisioned analytics jobs. Live, billable resources with no work behind them — and no “recommendation type” to put them in.

We made it concrete and asked AWS directly: “what idle resources should I clean up?” It returned only unused disk volumes — no idle load balancers, no gateways. And one volume it told us to delete was still attached to a running instance; it even said “detach it first.” Confidently telling you to delete something in use is the fastest way to lose trust. Sherpa had already excluded that one and surfaced the genuinely orphaned resources.

The Savings Plan tell

This is my favorite finding. A Savings Plan is a one-to-three-year commitment. AWS’s agent sizes that commitment to your spend today — waste included. Sherpa subtracts the optimizations first, then sizes it to what your bill will actually be once you’ve cleaned up.

Here’s the tell: AWS’s recommended commitment came back larger than Sherpa’s. Not because it found more savings — because it was about to lock you into a multi-year bet on idle and oversized resources you’re deleting next week. A bigger Savings Plan number isn’t a better one. Optimize, then commit.

What a cost agent simply can’t reach

Some questions don’t live in cost data at all, and that’s where the gap stops being about counting:

• Marketplace agreements. We asked; AWS’s agent said, plainly, it can’t — “not exposed via Cost Explorer, Cost Optimization Hub, or other billing APIs.” Sherpa inventories every agreement, vendor, status, and renewal.
• Private-pricing and enterprise-discount commitments. Sherpa tracks the burndown and warns you when you’re on pace to miss your commitment — months early, while you can still act. A cost agent has no view of it.
• Cost that’s also security. Sherpa connected an idle load balancer to the exposed workload still sitting behind it. One “delete this idle thing” turned out to be the single highest-value security fix. AWS’s agent reads cost data — it can’t make that connection, ever.

The bill AWS doesn’t mention

The agent is free in preview. Running it is not. It’s a thin layer over services that meter: every question hits the Cost Explorer API at a penny a call, and the good rightsizing needs Compute Optimizer’s paid metrics tier, billed per resource. Turn on the advanced features across a real fleet and you’re paying a few hundred to a few thousand dollars a month to power it — for a picture that still misses the idle load balancer. Free agent, metered everything.

Credit where it’s due

AWS is genuinely strong at the AWS-native conveniences it owns the data for — cost forecasting, presentation-ready scheduled reports, reserved-instance math, and a polished anomaly flow that files a Jira or Slack ticket when costs spike. Real features; all are on Sherpa’s roadmap too.

The point

A FinOps agent that reasons over cost data is useful, and AWS validating that is the best thing to happen to this space all year. But cost is one domain. The waste that hurts most isn’t in it, the commitment risk isn’t in it, the security exposure isn’t in it — and neither is the fix. Sherpa reads across cost, security, observability, and inventory in one view, sizes commitments on your real spend, and then proposes the exact change.

We ran this on our own environment, in the open. The interesting version is running it on yours.

FAQ

What is the AWS FinOps Agent?

It’s a public-preview AWS service (announced June 2026) that answers cloud cost questions in plain language. It’s a natural-language layer over AWS’s existing billing services — Cost Explorer, Cost Optimization Hub, Compute Optimizer, Cost Anomaly Detection, and CloudTrail — and it can investigate anomalies and schedule recurring spend reports.

Does the AWS FinOps Agent cost anything?

The agent itself is free during preview, but the services it queries are metered. Cost Explorer API calls bill per request, and the higher-quality rightsizing recommendations rely on Compute Optimizer’s paid enhanced-metrics tier, billed per resource. Across a real fleet, powering the agent can run from a few hundred to a few thousand dollars a month.

What can’t the AWS FinOps Agent do?

Because it reads only billing and cost data, it can’t surface waste that has no cost-recommendation type — idle load balancers, underused NAT gateways, or unassociated IP addresses. It can’t read AWS Marketplace agreements, track private-pricing or enterprise-discount commitments, or connect a cost finding to the security exposure behind it.

How is Sherpa different from the AWS FinOps Agent?

Where AWS has a recommendation, Sherpa matches it — often to the dollar. Beyond that, Sherpa reads across cost, security, observability, and inventory in one connected view: it finds idle infrastructure AWS has no category for, sizes Savings Plan commitments on your optimized spend instead of your current waste, tracks Marketplace and private-pricing commitments, and proposes the exact fix to apply with your approval.