Optimize: A Spend Picture With Evidence
The bill jumped mid-month. Finance noticed after the invoice closed. By then the cluster that caused it had been scaled down, the engineer who scaled it up was on holiday, and the explanation lived in a Slack thread nobody could find.
Most cost tools would have shown you the jump: a chart with a tall bar. They wouldn't tell you why, whether it was intentional, or which of the recommendations on the next tab applied to it.
That's the gap we built Optimize for. It's one of the modules built on Frontier Investigation Agents, and its promise is short: your cloud pays only for what you use. Getting there takes evidence, not a prettier dashboard.
Why the usual approach breaks
Cost tools are good at aggregation. They group spend by service, tag and account, draw the trend, and add a list of rule-based recommendations: idle instance, unattached volume, low commitment coverage.
Three things go wrong with that list.
Rules can't tell justified spend from waste. A large instance at low CPU might be a memory-bound cache doing exactly its job.
Recommendations overlap. A rightsizing suggestion and a commitment suggestion can both claim the same compute spend, and the headline adds them anyway.
A chart can't answer a follow-up. The moment you ask why something went up, you leave the tool and start opening consoles.
What we built
Optimize turns saved billing data into a ranked list of cost-saving actions for AWS, GCP and Azure. Each selected cloud connection gets an Overview, Cost Explorer, Findings, Anomalies and Reports.
The important word is saved. Optimize keeps a stored picture of your billing evidence. Cost Explorer questions, anomaly investigations and reports all start from it, scoped to the account and the evidence in front of you.
01Connect
Billing read access
An AWS, GCP or Azure connection that can read billing data.
02Scan
Read-only
Billing evidence, Cost Health, and an agent checking every billed service.
03Findings
Ranked actions
Each with an estimate, effort, risk, reversibility and an evidence badge.
04Decide
Your team
Status and disposition on the board, the discussion on the card.
05Validate
The next bill
Potential savings is an estimate. The provider bill is the proof.
How it works
Billing evidence comes from the provider. AWS uses Cost Explorer, in USD. GCP uses the Cloud Billing Standard export in BigQuery, converted to USD on import. Azure uses Cost Management and keeps the billing scope's currency.
The first scan is read-only. Optimize reads billing evidence, calculates Cost Health (a score of how efficiently the account spends), and asks an agent to check every billed service with read-only cloud commands. The output is findings: individual savings opportunities, each with an estimate, effort, risk and reversibility.
A daily refresh keeps it current. Once the first scan has produced cost data, a Daily refresh switch pulls the latest provider billing data once a day. Turn it off and updates pause, while the saved evidence stays readable.
Questions start from the data. Cost Explorer has an Ask CloudThinker section. Your question opens a conversation already scoped to the selected account, the 6- or 12-month range and any service filter. From an anomaly, an investigation opens a fresh conversation with that anomaly's evidence and the exact cloud connection in scope.
Every finding says how it knows
The piece that does the most work is the evidence badge. Each finding carries one.
01Verified
Read the live resource
The agent read the resource itself and confirmed every value.
02Inferred
Billing and config
Derived from billing and configuration data, without reading the resource.
03Assumed
Partly unchecked
Part of the finding could not be checked. Confirm it before you act.
The badge qualifies a finding. It doesn't rank it: an inferred finding can still be the largest saving on the account. But only verified, low-effort, reversible findings become Quick wins. A plausible opportunity that needs more context is Needs review. Spend the scan checked and found justified is No action, and it stays out of potential savings.
That third class matters more than it looks. A list that only ever says "cut this" trains people to ignore it. "We checked this and it's fine" is a finding too.
Design decisions
Data sizes the problem, the agent explains it. An anomaly's estimated impact and severity always come from saved billing data, never from an agent's judgment. The agent can link findings as the likely cause. It doesn't get to decide how big the problem is.
01Saved billing data
Sizes the problem
- Spend and forecastCurrent month so far, and where it may finish
- Anomaly impactActual cost above the baseline, and severity
- Cost HealthHow efficiently the account spends
Never the agent’s judgment
02The agent
Explains it
- FindingsRead-only checks of every billed service
- InvestigationsLive cost breakdowns and change history
- Explained byFindings linked as the likely cause
Says so when it cannot tell
03Your team
Decides
- StatusPending, In Progress, Implemented, Ignored
- DispositionWhy a finding left the queue
- DiscussionKept next to the evidence
Owns what changes
Potential savings is an upper bound, and we say so. The headline sums open Quick win and Needs review estimates. Estimates can overlap, so the total isn't a guaranteed bill reduction and shouldn't be subtracted from the forecast. Optimize doesn't present it as ROI or realized savings. If you need financial proof, check the next provider bill after you implement.
Missing is not zero. Provider billing data arrives late. When a refresh is delayed or blocked, Optimize shows the last successful evidence and how fresh it is. A Cost Health score that isn't ready reads Preparing, not zero. Reports name freshness, partial coverage and missing connections instead of filling the gaps quietly.
Unlike currencies never add up. A report across all accounts keeps each account in its own section, so an Azure scope billed in another currency is never summed with USD.
Decisions stay next to the evidence. Findings move through Pending, In Progress, Implemented and Ignored. A disposition records why a finding left the queue without erasing it: Expected, Snoozed or Disputed. Anomalies get their own: Unexpected change, Expected change or Insignificant. The Discussion tab keeps the argument on the card.
Anomalies are episodes, not alerts. One card can cover a multi-day increase. The daily scan extends it while the increase continues, and retracts it with a recorded reason when the supporting data no longer holds. Findings an agent linked as the cause show up in the anomaly's Explained by list.
In production
F88 reduced AWS spend by 30% with continuous cost and resource optimization, while automating 80% of its daily operations.
Honest limits
It isn't real-time. Anomaly cost is delayed billing evidence, not monitoring data. Optimize sees a spike after the provider bills it.
It doesn't change your infrastructure. The scan is read-only. Implementing a finding is your team's call, and the board tracks it.
It won't invent a cause. If credentials, provider access or the available evidence can't establish why spend moved, the investigation says so. That's less satisfying than a confident guess, and more useful.
Reports stay inside the evidence. They don't claim realized ROI, budget variance or purchase advice.
Try it
Connect an AWS, GCP or Azure account with billing read access and let the first scan finish. Then ask:
"Why did compute spend go up last month compared with the month before?"
"Which open findings on this account are verified, and which still need someone to confirm them?"
"Generate a savings opportunities report for all accounts and point out any account with stale billing data."
Start a trial or book a walkthrough.