Chatbot Knowledge Base Audit Template: Copy, Grade, and Launch

Start every audit with a header that names the chatbot, the source systems reviewed, the number of customer intents audited, the date, the reviewer, and the next review trigger. This makes the audit a living document rather than a one-time check.

Scope the audit to the intents the chatbot will answer first. Do not try to audit every article in the knowledge base. Focus on the top launch intents and the highest-risk exceptions.

Coverage measures whether each top customer intent has a focused, current source. Score coverage by intent, not by article count. A knowledge base with 500 articles but no answer for the top refund question scores low.

Score bands: 0-39 missing critical coverage; 40-59 partial coverage with gaps; 60-79 strong coverage with minor gaps; 80-100 complete coverage for launch intents.

Freshness measures whether sources were reviewed after the latest product, pricing, policy, or workflow change. An article can be old but fresh if it was verified recently. An article can be new but stale if it skipped the latest update.

Weight high-risk topics higher. A stale article about return policy is more dangerous than a stale article about company history.

Conflict measures whether help articles, macros, SOPs, and policies agree. The AI can retrieve multiple sources and blend them. If they contradict, the chatbot may produce a confident wrong answer.

Score conflict by intent. One clear canonical source scores high. Two or more conflicting sources score low and should block the intent until resolved.

Escalation clarity measures whether sources tell the chatbot when to stop and hand off. Articles that say contact support without specifying why are weak. Articles that name the exact conditions requiring human review are strong.

Escalation is not failure. It is the correct behavior for complaints, legal threats, privacy requests, security issues, regulated advice, account changes, and high-cost exceptions.

The audit should produce a launch recommendation based on the overall score. Use score bands to decide whether the chatbot is ready for broad automation, narrow pilot, or source-fix first.

Export the launch scope as a shared artifact: approved intents, restricted intents, blocked intents, source-fix-needed intents, and human-only intents. This becomes the configuration boundary for the chatbot platform.

Every blocked or source-fix-needed intent should become a work item. The backlog should include the exact source to fix, the owner, the deadline, the customer-safe wording, and the retest prompt. Sort by launch impact.

The retest plan should specify which customer phrasing to rerun after each fix. Without retest prompts, teams often fix the source but forget to verify that the chatbot now answers correctly.

The final audit should be exportable as a report that support ops, legal, compliance, product, and vendor admins can all read and act on. Include the score bands, launch scope, conflict table, source-fix backlog, retest plan, and reviewer signatures.

The report becomes the compliance record, the launch boundary, and the post-launch monitoring baseline. Update it when sources change, wrong answers appear, or the chatbot scope expands.