How do you keep an AI's knowledge base current?

Keep an AI's knowledge base current by naming a single owner for freshness, automating diffs between the source policy and the article that quotes it, sampling AI conversations weekly for outdated answers, and treating any policy change as a content task before it is a release. Without those four, knowledge decays predictably from launch onward, and the AI quietly starts giving stale answers.

On day one, the help centre an AI agent reads from matches the policy. On day ninety, the policy has changed three times, the article reflecting it has not, and the AI is now confidently quoting the old version to customers. The AI did not get worse. Its source did. Maintenance of that source is the work that decides whether the AI is useful in month six.

What people in the field are saying

kdschemin's "Reliability is the product" argues that in AI customer service, the reliability of the underlying data is the product. Trade press picks up the same theme as "knowledge ops" or "AI content operations."

Why does the knowledge base decay?

Because the people who change the policy and the people who write the article that reflects it are not the same. Pricing changes; the pricing page in the help centre may or may not. A return-window policy moves from 30 days to 14; the article quoting 30 days is still live. Nobody is consciously deceiving anyone; the link between change and content is just absent.

What does a current knowledge base look like?

Three properties. First, every article has a named owner, not a team. Second, every policy or product change automatically opens a content task against the article that quotes it. Third, the AI's audit log surfaces any answer it gave that quoted an article more than a fixed number of days old, so freshness has a measurable signal.

How do you catch the staleness the system misses?

Sample AI conversations every week and read them. Look for answers that sound right but feel out of date: prices, hours, return windows, contact methods. Pull the article the AI quoted. Check it against the live policy. If the two disagree, you have found one. Multiply by the volume of contacts going through the AI, and the cost of doing this once a week pays for itself many times over.

What is the smallest first step?

Name an owner for the top ten help-centre articles by AI traffic, today. They review each article against the live policy, monthly. That is enough to catch most of the damage. The rest of the discipline (automation, audit logs, full coverage) can grow from there.