AWS FinOps Agent is in preview.
I ran it on my real account.
I've been experimenting with AWS FinOps Agent since it dropped in preview. Set up an agent in under 5 minutes, ran a task to analyse my S3 costs, set up an automation for weekly HTML reports, and triggered the PPT generation. The agent found 4 cost anomalies, flagged Transfer Family as the surprise $27.61 spike, and handed me a 6-slide executive deck — all without me touching Cost Explorer manually. Here's everything.
- What is AWS FinOps Agent
- 3 trigger types: Run once, Schedule, Event-based
- Running a task: S3 Storage Cost analysis
- The artifacts it generated (HTML + PPTX)
- Setting up a weekly automation
- The executive PPT — real data, 6 slides
- The 4 anomalies it found (2 still active)
- Savings recommendations
- Context files — teach the agent your environment
- Honest takeaways
What is AWS FinOps Agent
AWS FinOps Agent is a managed AI agent — built on Bedrock — with direct, IAM-governed access to your AWS cost data. You describe what you want in plain English and it does the work: querying Cost Explorer, correlating with CloudTrail, building reports, posting findings to Slack or Jira. No dashboards. No manual queries.
It's live under Cloud Financial Management in the AWS Console, in public preview. Setup takes under 5 minutes — 5 guided steps, auto-created IAM roles, and you're running.
3 trigger types — Run once, Schedule, Event-based
When I first opened the web app I wasn't sure what separated Tasks from Automations. Both use identical forms with the same instruction box and the same When to run options. There are three trigger types, and both Tasks and Automations support all three:
Both Tasks and Automations share the same instruction box and the same "When to run" options. There are three trigger types — and both Tasks and Automations can use any of them:
So Tasks vs Automations isn't really about trigger type — it's about intent. Tasks are investigations you delegate. Automations are the always-on cadence your team runs on. Both can use all three triggers, including event-based.
Running a task: S3 Storage Cost analysis
First thing I ran — typed "Analyse the S3 Storage Costs" in the instructions field and hit go. The agent picked up the cost-analysis skill, called Cost Explorer in parallel for 6 months of S3 data, and came back in about 2 minutes.
Created: 9 June 2026 at 22:16 (UTC+5:30)
Started: 9 June 2026 at 22:17 (UTC+5:30)
Completed: 9 June 2026 at 22:19 (UTC+5:30)
Priority: Normal
S3 Storage Cost Analysis — Findings
Period: Dec 2025 – Jun 9, 2026 | Metric: UnblendedCost (USD)
Key Finding
S3 costs are effectively negligible. Total spend across 6 months
is $0.0000041 USD — operating almost entirely within the AWS Free Tier.
Ran for 2 minutes (22:17 → 22:19), pulled 6 months of S3 data, and surfaced the key finding: total S3 spend across the entire period was $0.0000041. Sub-millionth of a dollar. But what I appreciated — it still broke it down by month, identified Data Transfer as 91% of all S3 spend, and flagged May as the peak. The detail was there regardless of spend scale.
| Month | Total Cost (USD) | Notes |
|---|---|---|
| Dec 2025 | $0.0000000000 | $0.00 — baseline |
| Jan 2026 | $0.0000000000 | $0.00 |
| Feb 2026 | $0.0000000000 | $0.00 |
| Mar 2026 | $0.0000000553 | First activity |
| Apr 2026 | $0.0000006624 | Growing |
| May 2026 | $0.0000032700 | ← peak month |
| Jun 2026 (MTD) | $0.0000000662 | Month-to-date |
Spend only started appearing in March 2026 and grew month-over-month — sub-cent the entire time. The agent noted this trend explicitly, which is exactly what you'd want from a FinOps analyst reviewing the data.
The artifacts it generated — HTML report + PPTX
The task produced a structured result in the task view and a proper HTML artifact saved to Artifacts. The PPT came from a separate prompt asking for an executive summary. Both landed as downloadable files with unique IDs, timestamps, and source task links — traceable back to the exact run that generated them.
The HTML report has proper stat cards, a spend trend chart, and a free-tier status callout with a green tick. Opened it in the browser — I could send it directly to a stakeholder without touching it. The banner reads "Generated by AWS FinOps Agent" which is useful for traceability. The PPT I'll get to in the next section.
Setting up a weekly automation
After the task, I set up a recurring Weekly Cost Report automation — every Thursday 3:50 AM IST, generates a full HTML cost trends report automatically.
Every Thursday the agent wakes up, pulls the latest cost data, and drops an HTML report into Artifacts. No human involved. With Slack connected it would post the link there too. Timezone is fully configurable — IST is there out of the box. One thing to note: time is converted to UTC before scheduling and may shift by 1 hour during daylight saving transitions — the console flags this.
The executive PPT — real data, 6 slides
Most impressive output. Prompt was: "Summarize cost trends and savings opportunities in an executive-ready report in PPT." The agent powered up cost-analysis + create-presentation, called 4 tools in parallel, read design references, and handed back aws-cost-executive-report-2026-06-09-r3.pptx (380.2 KB).
It narrated what it built before sending the file — which I found genuinely useful:
Slide Content
─────────────────────────────────────────────────────────
1 Cover — Title, date, data source
2 Executive Summary — YTD KPIs, monthly trend table & bar chart
3 May 2026 Breakdown — Service table with MoM deltas + spend distribution donut
4 Cost Anomalies — 4 detected anomalies (2 still active: EC2 & VPC, Jun 3–7)
5 June Forecast — $20.72 projected, 37% below May; MTD pace analysis
6 Savings Opportunities — 3 prioritised recommendations + 4-step action plan
Key story in the data:
• May was the peak month at $32.87 — driven almost entirely by AWS Transfer Family ($27.61, 84% of total)
• June is forecast to drop to $20.72, but MTD run rate (~$2.15/day) implies higher — worth watching
• Two anomalies remain unresolved (EC2 Compute + VPC, Jun 3–7, combined $3.88 impact)
• No rightsizing recommendations returned by Cost Optimization Hub — enabling Compute Optimizer opt-in is the suggested next step
That narrative summary before handing over the file is genuinely useful. You know exactly what's in the deck before you open it, and the "key story in the data" section is the kind of synthesis a good FinOps analyst writes — not just raw numbers.
Slide-by-slide: what the agent actually built
I opened the PPTX. Here's what's in each slide with the actual numbers from my account:
| # | Slide | Key data points |
|---|---|---|
| 01 | Cover | AWS Cost Trends & Savings Opportunities · Executive Summary | June 2026 · Source: AWS Cost Explorer | UnblendedCost | Jan–Jun 2026 · Generated by AWS FinOps Agent |
| 02 | Executive Summary | May 2026 Total: $32.87 (highest month YTD) · MoM change Apr→May: +$28.14 (+594.8%) · Jun 2026 forecast: $20.72 · Active anomalies: 2 (EC2 & VPC Jun 3–7) · Full monthly trend table with bar chart |
| 03 | May 2026 Breakdown | 13 active services · Daily average $1.06 · Top: Transfer Family $27.61 (84%) · IAM Access Analyzer $1.80 (5.5%) · EC2-Other $1.57 (4.8%) · Route 53 $1.03 (3.1%) · Spend distribution donut chart |
| 04 | Cost Anomalies | 4 total anomalies · 2 still active · $4.46 total impact · EC2 Compute Jun 3–7: $2.60 (Active) · Amazon VPC Jun 3–7: $1.28 (Active) · EBS Apr 14: $0.35 (Resolved) · EC2 Compute Apr 8: $0.23 (Resolved) |
| 05 | June Forecast | Jun MTD (9 days): $19.34 · Full-month forecast: $20.72 (range $19.20–$22.24, 80% CI) · vs May: -37.0% · BUT: $2.15/day run rate implies ~$64 full-month — monitoring flag added |
| 06 | Savings Opportunities | 3 prioritised recommendations · Investigate Transfer Family ($27.61 impact) · Resolve Active Anomalies ($3.88 combined) · Monitor IAM Access Analyzer ($1.80) · 4-step action plan |
Slide 5's observation is the one I keep coming back to — the agent didn't just report the AWS forecast of $20.72. It flagged that the current $2.15/day MTD run rate implies a ~$64 full-month projection, well above the forecast. That's active reasoning on top of raw API data, not a widget.
The 4 anomalies it found (2 still active)
The agent pulled from Cost Anomaly Detection and surfaced all 4 anomalies from the Apr 1 – Jun 9 window. Two still active:
| Service | Date Range | Actual Spend | Expected | Status |
|---|---|---|---|---|
| EC2 Compute | Jun 3–7, 2026 | $2.60 | $0.00 | Active |
| Amazon VPC | Jun 3–7, 2026 | $1.28 | $0.00 | Active |
| Amazon EBS | Apr 14, 2026 | $0.35 | $0.02 | Resolved |
| EC2 Compute | Apr 8, 2026 | $0.23 | $0.02 | Resolved |
Both active anomalies on EC2 and VPC had $0.00 expected spend — these services were essentially idle before Jun 3. That's a real signal. The agent correctly flagged: review in Cost Anomaly Detection console and confirm expected vs unplanned.
Savings recommendations
Slide 6 is the "what do I do next?" slide. Three prioritised recommendations with actual dollar impact:
The agent also produced a 4-step action plan at the bottom of the slide:
One thing I appreciated: when the Cost Optimization Hub had no rightsizing recommendations, the agent didn't just leave that section blank. It wrote: "This may indicate resources are already optimised, or that Compute Optimizer opt-in is pending." That's the difference between a tool and an analyst.
Context files — teach the agent your environment
This is the feature I hadn't fully appreciated until I saw the Console panel. Inside the agent's web app there's a Context files section — and right now mine reads: "Context files (0) — Help FinOps Agent learn your environment by uploading relevant documents."
The idea is simple but powerful: right now the agent knows my AWS cost data, but it doesn't know my org. It can't tell me "this EC2 instance belongs to the Platform team" or "this S3 spike is expected because we run month-end batch jobs." Context files close that gap.
What's worth uploading here:
Owner, Project, CostCenter) — so the agent can flag untagged resources and map spend to teams.Without context files, the agent gives technically accurate but operationally generic answers. With them, the same anomaly response goes from "EC2 spend increased by $2.60" to "The Platform team's ECS cluster in us-east-1 has unexpected EC2 spend — this should be reviewed by the on-call engineer." That's the unlock for enterprise use.
Honest takeaways
I went in expecting another "AI-powered" dashboard wrapper. I came out genuinely impressed. Here's what actually stuck with me:
- Three triggers, one interface. Run once for ad-hoc investigations, schedule for recurring cadence, and event-based for reactive anomaly response. Same instruction box for all three. The event-based trigger is the one that eliminates the human router between a Cost Anomaly Detection alert and the engineer who owns the resource.
- The PPT quality surprised me. Slide 5 flagging that the MTD run rate implies ~$64 vs the $20.72 forecast isn't something you get from a Cost Explorer widget. It's synthesis, and it was correct.
- Context files are the enterprise unlock. Without org-specific context, answers are accurate but generic. Upload team ownership and tagging standards and the agent starts giving you operationally actionable findings, not just data.
- The HTML artifact is shareable as-is. Proper stat cards, trend chart, free-tier callout. I could email it to a stakeholder without touching it.