Back to Blog Home

How to Keep Your Company Brain from Going Stale: Automated Content Decay Prevention

Person in a sweater typing on a laptop at a wooden desk

Do not flag pages by age. A page nobody has touched in a year can be the most correct page you have, and an age rule buries owners in false alarms. Flag a page when the thing it describes changes. Review first the pages your assistant actually cites, so a page that appears in fifty answers and was confirmed a year ago comes before one that nobody has opened. Keep AI for the narrow jobs a rule cannot do: spotting a contradiction with a newer source and drafting the update for the owner to approve. For every page type, decide four things: who owns the page, what flags it, how long it may go without a check, and what the assistant says when the check is overdue.

A decay checklist by page type

Write those four decisions where the owner can see them. Keep "last edited" and "last confirmed" as separate dates. A review ends in a re-confirm or an edit, and a page confirmed yesterday is current even if nobody has edited it for two years.

Page typeWhat flags itOwnerBack-stop interval
How-to for a system (deploy, access, setup)A pull request or release that changes the systemThe team that runs the system6 months
Policy (expenses, leave, security)The source policy or its effective date changesThe policy owner in People, Finance or Security12 months, 3 months for security and payroll
Reference (org chart, vendor list, glossary)The system of record changes, so link to the record instead of copying itWhoever owns the record12 months
Customer-facing playbookA release or a pricing changeThe product or support leadEvery release
Decision record or historyNothingThe author's teamNever. Label it historical, with its date, so it is cited as history and not as current practice

The intervals are starting points. Tune each one by what reviewers do with it: if a page type is re-confirmed unchanged almost every time, lengthen its interval, and if most reviews end in an edit, shorten it and look for the trigger that should have fired earlier. Assign pages to a team and not to a person, so a departure changes a team's membership and not hundreds of page owners. In Kipwise, content review works this way: you name a reviewer and a review interval for a page or a folder, reviewers are reminded, and a page shows when it was last approved.

Flag by change first, use AI second

Most staleness has a cause you can name: a release, a pull request, a policy change, an owner who left. Catch those with rules, which are cheap, explainable and exact about whether a pull request touched a path. For documentation about code, GitHub already has the mechanism. Code owners are automatically requested for review when someone opens a pull request that modifies code they own, though not on draft pull requests, so make the page's owning team a co-owner of the code the page describes:

# .github/CODEOWNERS
/deploy/    @acme/platform-team @acme/runbook-owners

Two constraints decide whether this works for you. Code owners must have write permission on the repository, so a support lead who owns the page may need to sit in a team that has it. And the last matching pattern wins, so list every owner on the same line. For what a CODEOWNERS file cannot express, a GitHub Actions workflow can run when a pull request changes certain file paths (see the path filters), then comment on it or open an issue for the owner.

AI earns its place in the gaps a rule cannot see:

  • comparing a page's volatile details, such as versions, dates, names, prices and URLs, with the source of truth and listing the differences;
  • finding two pages that answer the same question differently;
  • drafting the update from the changelog or the diff, for the owner to approve;
  • grouping the questions the assistant could not answer into missing pages.

It should not decide on its own that a page is stale, edit a page, or read channels in the background to harvest corrections. Use the questions people ask the assistant itself, which they know are being processed. Extraction services such as Google's Document AI extract, classify and split documents, with mail-room, mortgage and procurement paperwork as the typical uses, and they know nothing about staleness. They help only when your sources are scanned PDFs, where OCR turns them into text a check can read.

Review what the assistant cites first

Reviewer time is the scarce resource, and oldest-first spends it on pages nobody reads. We recommend ordering the queue by how often the assistant cited a page since its last confirmation. When a pull request or release has changed the thing a page describes since the page was last confirmed, have the assistant add that caveat to the answer instead of answering as if nothing happened.

A worked example: the deploy runbook

  1. The runbook's owners, @acme/runbook-owners, are co-owners of /deploy/, so a pull request that changes the pipeline requests their review.
  2. The owner opens the runbook from the pull request. If it still holds, one click re-confirms it and edits nothing. If not, they edit it, and the edit confirms it.
  3. If the owner does nothing, the page keeps its old confirmed date. The assistant cites it with the caveat that the pipeline changed after the last confirmation, and the page moves up the review queue because it is cited often.
  4. After six months with no trigger, the back-stop interval sends the owner one question in Slack, "Is the deploy runbook still accurate?", with two answers: still correct, or needs changes.

Measure it, and keep it quiet

Judge the system by three numbers: the share of cited pages confirmed within their interval, the share of reviews that end in an edit for each page type, and the wrong-answer reports traced to a stale page. Aim for a short gap between a change and the corrected page. Flag only what a trigger or a comparison supports, because alerts on every minor inconsistency teach owners to ignore them, and give every flag a one-click "still correct" so a false alarm costs seconds. Keep a human approval on edits to critical pages, and take reviewer complaints about a wrong flag as a bug in the rule.

Review prompts go only to a page's owners. A restricted page's title or excerpt must never appear in a prompt, a digest or an answer shown to someone who cannot open it. When a page loses its owner because of a reorganization, route it to the owner's team or to one named knowledge manager, never to nobody.

Start with one page type

Pick the page type your assistant cites most, usually the how-to pages for one system. Name the owning team, set the trigger and the back-stop interval, and run one review cycle before you add the next page type.

Topics

Related posts

Your team’s knowledge, answered in Slack

Add the Kipwise Agent. It answers from your pages with the source, and asks before it changes anything. 7-day trial, no card needed.