Guides: from AI claims to checked work

These guides turn a broad AI conversation into specific decisions about work, evidence, data, and accountability. They are editorial recommendations, with supporting sources and limitations stated in each article; they are not product test results or jurisdiction-specific legal advice.

New to the subject? Follow Start Here instead of reading by publication date.

• Choose the workflow before you choose the AI: Define a task, owner, and fallback before choosing a product.
• Ask what the system actually does: Separate extraction, generation, retrieval, and agents in a vendor explanation.
• Design a pilot that can produce a no: Write the acceptance and stop rules first.
• Count missing fields: Evaluate accuracy with a visible denominator.
• Measure the work left: Include preparation, review, and correction in the time comparison.

• A source link is not proof: Check whether the cited source supports the output.
• Give the reviewer a job: Define review responsibilities and authority.
• Before uploading a client document: Check the account and the data path.
• Give the assistant a permission boundary: Separate drafting from external actions.
• The document is evidence, not an instruction: Understand indirect prompt injection.

• Confidence is not approval: Use confidence scores without mistaking them for correctness.
• The checking should not end with the pilot: Monitor change, defects, and manual fallback.

Pair your reading with the working resources. The worksheets are deliberately blank: your team supplies the observations and decisions.