AI Vanguard matched-case cost checker v1 — 8 November 2026

Python 3.10+; standard library only. Reads a local CSV and prints JSON.
No API calls, network access, credentials, file modifications or model execution.

Save the Python file and synthetic-cases.csv in one folder. Run:
python check_workflow_cost.py synthetic-cases.csv --hourly 60
Use python3 or py instead of python if that is your local Python command.
Exit 0 means valid arithmetic input; it does NOT mean the agent passed a quality gate.
Exit 2 means invalid input. A workflow with zero accepted cases returns null cost per accepted case.
The complete ZIP includes the 12 regression tests. After extracting, run:
python -m unittest -v test_check_workflow_cost.py

Synthetic data is deliberately constructed to demonstrate review burden and rejected work.
It is not observed model output, client data, a representative sample or a benchmark.
At $60/hour the baseline costs $12.00 per accepted case, candidate-a $23.37,
and candidate-b $4.20. Candidate-a accepts only 3/5 cases; the others accept 5/5.

Replace every row with your own recorded cases. Required CSV headers:
case_id,workflow,work_minutes,review_minutes,rework_minutes,api_tool_usd,accepted,evidence_ref

Use the same case IDs for each workflow; include a workflow named baseline (or use --baseline NAME).
One row represents the COMPLETE work on one case in one workflow, including all retries.
Run separate files for repeated trials; this tool does not estimate statistical uncertainty.
work_minutes: human preparation, prompting and hands-on work, excluding review and rework.
review_minutes: human checking/approval time, excluding work and rework.
rework_minutes: human correction/escalation time not counted in the other two fields.
These three time columns must not overlap. Model waiting time is a separate cycle-time metric.
api_tool_usd: actual model/tool charges for the whole case, including failed calls and retries.
accepted: exactly yes or no, using criteria set before the run and a consistent observation window.
evidence_ref: a local ticket/trace/check reference. This script checks presence, not truth.
Keep sensitive ticket details locally; the CSV is never uploaded by this tool.

Formula: (all human minutes * hourly rate / 60 + all API/tool charges) / accepted cases.
Costs of rejected cases stay in the numerator. Acceptance is separately reported.
Do not select an option solely for low cost if correctness or safety is worse.

Important limits: IDs do not validate task mix or grading. Small samples do not establish reliability.
Add setup, subscriptions, infrastructure, training, downstream failures and uncertainty separately.
Use separate rows/runs or a consistent blended rate when staff costs differ; the command uses one rate.
Do not count saved labor as cash savings unless spending actually falls or capacity is usefully redeployed.
No license to act on systems is granted. This is measurement arithmetic, not incident response.

Source and article: https://aivanguard.tech/ai-agent-roi-framework-2026/#cost-checker
Model-selection context: https://aivanguard.tech/best-ai-models-2026/#ms-cost-proof
Code license: MIT. Copyright (c) 2026 AI Vanguard.
Permission is granted, free of charge, to use, copy, modify, merge, publish, distribute,
sublicense, and/or sell copies of this software, provided this notice is included.
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED,
INCLUDING MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.
THE AUTHORS SHALL NOT BE LIABLE FOR ANY CLAIM OR DAMAGES ARISING FROM THE SOFTWARE.