Building dependable AI support
How do you prepare a knowledge base for AI support?
Prepare AI support knowledge by writing clear answers to real customer questions, including the conditions under which each answer is true. Remove outdated or conflicting policies, keep account-specific details out of general articles, and test both the normal case and the exceptions before relying on the content.
What belongs in an article?
State who the answer applies to, the steps they can take, and any restriction that changes the outcome. Define product terms customers may use differently. A short article with explicit prerequisites is often easier to maintain than a long document that mixes several policies without clear boundaries.
How do you handle exceptions?
Keep the exception beside the rule it modifies. If a feature is available only on a particular plan or in a specific country, say so in the same explanation. Otherwise a general statement can be retrieved without the qualification that makes it accurate.
How do you keep knowledge useful?
Assign someone to review policy changes and use real conversations to find missing details. When two articles disagree, resolve the source conflict instead of adding another instruction asking the model to guess which one wins. Re-run questions that previously produced incorrect answers.
Practical checklist
- State prerequisites and exclusions explicitly.
- Remove superseded or contradictory information.
- Separate general policy from private customer records.
- Test one common question and one exception per article.
Worked example
Instead of “You can edit a campaign,” write “An organizer can edit a draft campaign. After the campaign is locked, contact support; another teammate’s editing permission does not unlock it.” Test a question about a locked campaign to check that the limitation survives.
How Fire Away fits
Fire Away uses knowledge articles and approved website drafts for support answers. Agents can also propose knowledge from a conversation, review the draft, and approve it rather than silently turning every chat into policy.