Prepare useful sources
Choose the correct bot and inspect its current sources before adding content. Use authoritative documents, a relevant website, concise text or Q&A pairs. Remove obsolete or contradictory material and decide which facts the bot should be able to answer.
Keep credentials and unnecessary personal data out of training content. The knowledge upload playbook explains source selection and the existing classic upload screens.
Onboarding when available
If onboarding is offered, provide the website or supported attachment requested by the app. Read the proposed bot setup, selected content and sample questions. Each consequential action needs its own review card and an explicit click on its approval button. Use Apply on a hosted workflow review or Approve and its send, publish or delete variants on a standard review. An ordinary chat yes is not approval.
Bot creation, source changes, training and paid tests are separate decisions. Informational plans have no blanket Apply. If onboarding is unavailable, use Classic to create the bot and add supported sources; do not assume that a chat request has started an automated workflow.
Refresh when available
For an enabled refresh workflow, identify the existing bot and website. Inspect the proposed changes and selected pages.
Pages of unknown freshness are not refreshed unless you include them. Where Edit is offered, it can include only listed pages of unknown freshness. It cannot remove new or changed pages. If the fixed selection is wrong, do not Apply it: Dismiss stops that refresh, or request a revised proposal. Apply only the exact reviewed effect. Review later retraining and testing separately.
After processing, check the source status and test results. A completed crawl or upload does not establish that every answer is accurate. It does not prove that an expert knowledge base has finished updating. If the task is interrupted or synchronization is incomplete, check its evidence before deciding what to do next.
Read test and citation evidence
Prepare representative questions with expected behavior: a correct fact, a question requiring a citation, and a question the bot should decline because the source does not support it. Where Atlas offers expert test questions, a reviewed test can run up to five questions. Provider cost may be unknown in advance.
A successful invocation means the expert returned an answer. A quality pass is a separate assessment. Read each answer, expected behavior, verdict and reasons. Missing evidence stays unknown; an answer without enough evidence is not a verified pass.
Check citations against the actual document and relevant passage. A link alone does not prove the answer is supported. See citation quality guidance. Record which sources and time period support your conclusion, and test again after a reviewed content change when appropriate.
If an earlier paid test may have run but its result is missing, follow unknown outcome guidance before requesting another invocation.