A privacy-first, read-only cost-leak scan. We detect duplicate and retry calls, batch opportunities, caching gaps — including caching that's enabled but expiring — and attribution gaps. Raw prompts don't need to leave your AWS account.
Read-only IAM role. No proxy, no SDK, nothing in your request path. We surface findings; you decide what to fix.
Cost Explorer and CUR tell you which team or model spent the money. They can't tell you that a stable prompt prefix is being re-sent thousands of times with caching off, that a nightly job is paying on-demand rates for work a batch queue would halve, or that retries are quietly double-charging you.
Agentic Bedrock workloads make this worse: fan-out, retry loops, and redundant context are invisible in an aggregate bill and tedious to find by hand in raw invocation logs.
CostRoot reads those logs the way an engineer would — and shows you the specific, fixable patterns, ranked, each with the exact lever.
Read-only, from your invocation logs. Findings are surfaced for your review — nothing is changed in your account, and no dollar figure is claimed until you confirm you own the fix.
Sessionizes traffic per principal and flags repeated and retried invocations — retry loops and redundant context that re-pay for the same work. Findings are evidence-backed and ranked.
Spots scheduled, async-shaped workloads paying full on-demand rates where Batch inference applies
a flat 50% cut — no model or prompt changes required.
Finds large stable prefixes re-sent with cacheRead=0. On subagents Bedrock disables
caching by default; cache reads run roughly 90% cheaper than re-sending
the prefix.
The one most teams miss: caching is on, yet a stable prefix is being re-written every time the workload idles past the TTL — paying the write premium on cold re-caches. We catch the cache churn, not just the cache absence.
Where logs collapse spend to a single gateway or shared role, we mark what can't be traced to a team or task — so you know exactly where visibility ends, rather than trusting a number that quietly isn't there.
write permissions requested.
Remediation ships as a diff you review and
apply. We never act on your account.
One CloudFormation deploy creates a read-only role in your account, locked to an ExternalId. It can read logs and metrics. It can write nothing.
Logs are parsed for the four patterns above. Prompt bodies are hashed, not stored; your latency, reliability, and architecture stay exactly as they are.
A report of the fixable patterns we found, each with the lever. You confirm which fixes you own — only then does a finding become an actionable number.
A worked example of the type of waste CostRoot detects, built on realistic Bedrock usage patterns. It is not a real customer — the workload is generated, and the figures are illustrative of the method, not a promise about your account.
The design-partner program is for engineering and FinOps teams who run real Bedrock traffic and want a clear-eyed read on where it's wasteful — before committing to anything.
Multi-step agents, tool-calling, and retrieval pipelines where fan-out and retries add up.
Enough steady traffic that a few percent of waste is worth an engineer's afternoon to fix.
Teams who'd rather grant a read-only, auditable role than route production traffic through a proxy.
We're onboarding a small set of design partners. Tell us about your workload; if it's a fit, we'll run a read-only scan and walk you through what we find — no obligation, no sales pressure.
Qualification form. The scan is read-only and the findings are yours to keep, whether or not we work together.
The trust boundary, the exact read-only permissions, ExternalId handling, how prompt bodies are hashed rather than stored, and the access audit trail — laid out for a security review. Read the security brief.
No. Prompt bodies are hashed and reduced to token counts where the logs are parsed; raw text doesn't need to leave your AWS account, and we never persist it.
No. The role is read-only with zero write permissions. Every fix is a recommendation you review and apply yourself.
Neither. CostRoot reads invocation logs out of band. There's nothing in your request path, so your latency and reliability are untouched.
A ranked list of the fixable patterns we found, each with the specific lever — a cache point, a batch migration, a route change — and the evidence behind it.
Because an honest one only exists after you confirm you own each fix lever. From logs alone, a finding is a recoverable opportunity, not a guaranteed dollar amount.
We walk you through the findings. If they're useful and the fit is right, we can talk about an ongoing watch — but the scan stands on its own, no commitment required.