// AI Agents & LLM Ops

LLM Usage Cost Attribution Reconciler

Reconcile provider bills, usage events, model prices, retries, and customer allocation into deterministic AI cost findings.

Get it on Agensi Browse all skills $19 one-time

What problem does LLM Usage Cost Attribution Reconciler solve?

Provider bills, gateway logs, model-price tables, retry and cache events, and customer usage often disagree at invoice close or margin review.

Use it to

Reconcile an LLM provider invoice against gateway usage
Find retry and fallback costs without customer allocation
Catch model price-table drift before a margin review

What it returns

findings.csv with stable codes and evidence sources
review-manifest.csv with every reviewed record
result.json with READY, REVIEW, or BLOCK gate and SHA-256 receipt
report.md with a human-review handoff

A representative input and result

fixture-backed sample

input `billing_items[]` uses `billing_id`, `amount`, and optional `usage_event_id`; `usage_events[]` uses `event_id`, `provider`, `model`, retry evidence, and customer allocation; `price_table[]` uses `price_id`, `provider`, and `model`.

result A deterministic READY, REVIEW, or BLOCK result with stable finding codes, a complete review manifest, a human-readable report, and a SHA-256 receipt.

Access and approval boundaries

Terminal permission to run the bundled local Python reconciler.
Read access to one user-selected normalized UTF-8 JSON packet.
Write access only to one user-selected local output directory.
No browser, network, credential, account, environment-variable, messaging, payment, deletion, access-change, filing, or publication permission.

Known limitations

Owner-supplied price tables can be stale and must be refreshed by the responsible operator.
The package does not query providers, calculate a bill from raw tokens, or authenticate export completeness.
READY does not prove margin, billing accuracy, or profitability.

Questions

Does it connect to a live account or provider?

No. It reads one normalized local JSON packet and makes no network request.

Does it make the final decision or external change?

No. It produces evidence findings; the responsible owner makes every decision and action outside the package.

Is the result deterministic?

Yes. Stable finding codes, sorted records, and a SHA-256 receipt make repeated review inspectable.

What happens with malformed input?

Unreadable JSON, missing arrays, duplicate core identifiers, invalid timestamps, and invalid core numbers fail closed.

Does READY prove the underlying evidence is true?

No. READY means no automated finding appeared in the supplied normalized packet; source truth still requires human review.

LLM Usage Cost Attribution Reconciler keeps proof and approval boundaries visible.

The listing includes the tested package, realistic samples, declared permissions, and known limitations.

Get LLM Usage Cost Attribution Reconciler on Agensi