ai-ledger

Build-Time vs. Run-Time AI Record

A companion template for AI Helped Me Build It. AI Does Not Have to Run It.

Use one record per workflow. The point is to make AI’s role a design decision someone approved, not a default nobody noticed. It takes about fifteen minutes, and it answers the first question a reviewer will ask: “Where exactly is the AI in this?”

Copy this file, fill in the brackets, and keep it with the workflow (next to its README or Human Instructions).

1. The workflow

Field	Entry
Workflow name	[e.g. Monthly duplicate-payment test]
Business purpose	[one sentence]
Owner	[name / role]
Backup operator	[name / role]
Where it lives	[folder or repository path]
Version / commit in use	[e.g. v1.2 / a1b2c3d]

2. Build time: what AI did while the workflow was designed

Tick everything that applies and add a note.

AI helped to…	Yes / No	Notes
Explain the existing manual procedure
Draft the script, query, or formula
Explain the code line by line
Generate mock or synthetic test data
Suggest edge cases and test cases
Troubleshoot errors
Draft documentation

Tool and model used at build time: [e.g. Claude Code (model name), OpenAI Codex, GitHub Copilot agent mode, Copilot in Excel]

3. What data did the model see during the build?

Data type	Seen by the model?	Approved by
Synthetic / mock data only	[Yes / No]
Anonymized or sanitized production data	[Yes / No]
Production data	[Yes / No]
Column names and structure only (no values)	[Yes / No]

If anything other than synthetic data was used, note the approval and the data classification it fell under.

4. Run time: does AI have a role when the workflow runs?

Pick one.

No run-time AI. The approved script, query, or formula does all recurring work. AI is out of the production path.
Run-time AI for specific steps only. List them below.
Run-time AI coordinates the workflow (for example, an agent calling approved deterministic tools). List the tools and boundaries below.

Step	What AI does at run time	Why AI (not a rule)	Data it sees	Output reviewed by
[e.g. Draft variance commentary]	[Drafts questions for the account owner]	[Language interpretation]	[Variance summary only]	[Senior accountant]

5. Boundaries enforced at run time

Boundary	How it is enforced	Harness-enforced or instruction-only?
[e.g. Cannot read payroll folder]	[Claude Code `permissions.deny` rule]	[Harness-enforced]

See the Boundary Audit if you are not sure which column a boundary belongs in.

6. Approval of the design

Field	Entry
Design approved by	[name / role]
Date	[date]
Next review	[date, or “on any change to AI’s role”]
Linked release checklist	[link to completed Deterministic Script Release Checklist]

*PythonMuse LLC

www.pythonmuse.com. Educational purposes only; not professional advice. Content licensed under CC BY-NC-SA 4.0.*

This site is open source. Improve this page.