Resource / Practical, ungated guidance
Proposed AI-assisted delivery standard
A proposed baseline for an engagement, pending owner confirmation. This is not a claim that every practice is already operational.
Agree permissions before use
Proposed baseline: customer data and non-public materials must not enter generative AI systems without written approval. Document the permitted tools, information classes, retention settings, and people authorized to use them. Where approval is absent, use synthetic test data and public material only.
- Written permission and tool inventory
- Information classification
- Retention and access settings
- Synthetic fixtures for sensitive workflows
- Customer escalation path
Keep people accountable for engineering decisions
Proposed baseline: begin with an agreed specification, have a person review code and security boundaries, and test generated changes against acceptance criteria. AI output is a draft that can be rejected; it does not establish correctness or compliance.
- Specification and acceptance criteria
- Human code review
- Dependency and security review
- Automated and manual workflow tests
- Independent review for material changes
Make the user-facing boundary clear
Proposed baseline: any generated user-facing answer should identify itself, link to its evidence, respect permissions, and offer a human contact route. The Conduit website uses local semantic search over public content; it returns excerpts and page links and has no chatbot.
- Public or permitted source corpus
- Source links
- Clear limits and failure behavior
- No unsupported certification or pricing advice
- A route to a person