In 30 seconds

An AI workflow that answers from company knowledge is only as good as that knowledge. After launch, someone has to keep the sources current, because stale sources produce confident, wrong answers that look fine.

Why this comes up

At launch, the knowledge is fresh and the answers are good. Then the world moves: prices change, policies update, products are added, procedures shift. If nobody maintains the sources, the workflow keeps answering from the old version, and the errors are invisible because the answers still sound right.

This is one of the most common ways a workflow that worked in spring is quietly wrong by autumn.

Make maintenance a real, owned habit

Maintaining knowledge is not a one-time cleanup; it is ongoing operation. Building on how to keep knowledge current:

  • Every source the workflow uses has a named owner.
  • Each source has a review cadence matched to how fast it changes.
  • Reviews actually happen, tracked against the review date.
  • New and changed sources are added to the searchable set, and superseded ones removed.

The knowledge source review checklist gives you a repeatable check to run per source.

Notice when a source changes upstream

Sometimes a source is maintained by another team who do not know your workflow depends on it. A quiet change on their side becomes a wrong answer on yours. Where you can, arrange to be told when key sources change, or check them on a schedule. Treat a dependency on someone else's document as a real operational risk.

Connect maintenance to failures

Your failed case reviews are an early warning system for stale knowledge. A cluster of wrong answers on one topic often traces to a single outdated source. Use failures to find the sources that need attention.

A small example

An assistant kept quoting an old return period. The policy had been updated by the legal team, who did not know the assistant used it. Adding a quarterly review of key policies, plus a note to legal to flag changes, closed the gap, and the assistant's answers stayed correct through the next update.

What this does not solve

Maintaining sources keeps answers current; it does not guarantee the sources are correct, or that every upstream change is caught in time. It reduces stale-knowledge errors sharply, but combine it with monitoring and failure review to catch what slips through.

Good enough to move on when

  • Every source has an owner and a review cadence.
  • Reviews are tracked and actually happen.
  • Superseded sources are removed from the searchable set.
  • There is a way to learn when key upstream sources change.