Start with the questions customers already ask.
Choose a bounded area such as first-time setup or a particular feature. Gather the current answer for each question. Use your own support records where you have permission, and remove personal or account-specific information before turning them into reusable guidance.
Write down questions the agent should refer to a person. Keeping those boundaries explicit makes it easier to judge test conversations.
Choose a source for each answer.
Add an FAQ manually or select the GitHub document that contains the current explanation. Check instructions, links and version-specific details before approving it. Private repository access needs the repository on the workspace allowlist.
Prefer a clear answer with the conditions that matter. If a feature depends on a plan or product version, include that detail in the same source.
Prepare sales knowledge as its own collection.
Use approved sales notes for features, pricing and buying questions. Sales sources are separate from the support collection, so check that the sales agent has the public information it needs.
In the current pilot, website URL imports are disabled. Add website information manually and review it before approval.
Test the answers and keep approvals current.
Ask the same question in different ways. Include ambiguous questions and questions outside the sources. Read the transcripts and correct the underlying documentation when an answer is incomplete or confusing.
A GitHub sync requires reapproval before changed material becomes available. Repeat the relevant tests when a feature or plan changes. Approval is a publishing control, and answer quality still needs review.
Common questions
Does approving a source guarantee an accurate answer?
No. Approval controls which material the agent can use. Generated answers still need evaluation, including questions outside the source material and cases that should go to a person.
Can an AI agent update the knowledge base?
A trusted agent with a workspace management token can add and review sources through REST or MCP. Approval is a separate operation, so your process can keep a review step before publishing.