Your AI bill is not your cost per successful task.

Why a provider cost export is not an operational outcome report

Start with your provider’s billing source. OpenAI's monthly usage and cost export instructions explain how to obtain activity or detailed cost CSVs, including invoice-like groupings. The Claude Console cost and usage guide describes usage and cost breakdowns by model, workspace and API key. Those are legitimate ways to answer what your provider charged.

But provider cost records usually do not, on their own, prove whether a real task was completed for a customer. That mapping requires your own task or job identifier, a meaningful outcome state, and a trustworthy reconciliation between usage and business records. The downloadable CSV formats are provider-specific: this tiny tool does not import raw OpenAI or Claude billing exports directly. You must first create a safe task-level ledger that joins the billing evidence to your own authorized work identifiers.

Example: the cheaper-looking workflow that failed twice

In this completely fictional sample, one catalog job costs $0.18 on a failed attempt and then $0.25 on its successful retry. That job’s observed cost is $0.43, not $0.25. Another $0.41 catalog job has only a failed record. A third remains pending and costs $0.16. The ledger also includes an $0.08 row without a project label.

Run this with Load fictional sample. The result is $1.49 of supplied costs, $0.59 in failed-attempt rows, $0.16 in pending rows, and $0.08 without a project. Across four task IDs marked as reported successful, the observed total for those tasks is $0.92, or $0.23 each. That is a descriptive ledger calculation, not evidence that four customers were served or that the failed-attempt spend was avoidable. The job with both failed and success rows has an unknown final state: this simple ledger does not contain attempt timestamps, so the existence of a success row cannot prove the last attempt or eventual business outcome. The tool counts task IDs with a success row as an observation and separately flags mixed-success/failure IDs for investigation.

Compare alternatives before adding another dashboard

Provider billing + spreadsheet: Often the right answer if you have clean exports, stable task IDs and someone who can make a pivot table. Excel PivotTables can already summarize the same ledger with flexible groupings. Use the free native approach if it is enough.

Observability platform: Langfuse already supports detailed generation-level token and cost tracking, including ingested and inferred costs. If your team has instrumentation and ongoing needs, a tracing platform may outperform a manually pasted CSV snapshot.

This free snapshot: A low-setup, read-only investigation aid when you have a small authorized ledger and want to see failed-attempt spend, task-level reported outcomes and unattributed rows before deciding whether to invest in an integration. The test for real usefulness is simple: can a person do the same analysis faster or more accurately than with their provider export and spreadsheet?

What might justify a paid service?

Not a prettier chart. The possible paid work is controlled data extraction, matching model usage to existing job IDs, validating business receipts with the actual system owner, handling timing and retries, and recurring exception reports under an agreed scope. Those outcomes require explicit access, security review and a real buyer. Moruneva has not yet established demand or measured a cost saving from this experiment.

Related reliability question: when a workflow says success but the CRM record does not exist, use the Workflow Outcome Reconciler instead. That tool compares expected source events, executions and downstream receipts; it does not estimate your AI bill.