The answer was right, two years ago
A new manager asks the company AI assistant how many days of compassionate leave staff get. It answers clearly and cites a document. The document is "Staff Handbook v3 DRAFT (Jo's comments)" from before the policy changed. The current handbook is in a different folder with a slightly different name.
Once you look, you see it everywhere. Price lists from the year before last, a proposal template with the old company address, a process document for software you no longer use. The assistant is doing exactly what it was built to do: find the most relevant text. It has no way of knowing that relevance and correctness are not the same thing.
Why it keeps happening
Shared drives and SharePoint sites grow by copying. People save a version to edit, email an attachment that gets saved again, duplicate a folder for a new year. Over time the same document exists in many places with small differences, and the naming rarely says which is final. People cope because they know, or ask someone who knows. The AI cannot ask.
Most AI search setups index everything they can reach. Old drafts often match questions better than the final version, because drafts tend to be wordier and explain more. So the AI does not just sometimes pick an old version; it can prefer them.
What stale answers cost
| Stale source | What goes wrong |
|---|---|
| Superseded HR policy | Staff are told the wrong entitlement, which leads to disputes |
| Old price list | Quotes go out at prices you no longer charge |
| Outdated process document | New starters follow steps for a system you replaced |
| Draft contract template | Clauses that legal removed come back into use |
| Old org chart | Queries are sent to people who have left |
Trust drops fast. After a couple of confidently wrong answers, staff go back to asking colleagues, and the assistant becomes something people are told to use but do not.
How we get AI answering from the current version
- We scan the document stores the assistant uses (SharePoint, OneDrive, Google Drive, network shares) and group near-duplicate files by content similarity, not just file name.
- For each group we identify the likely current version using modified dates, approval metadata, location and naming, and produce a list for document owners to confirm rather than guessing on their behalf.
- We add a simple status to each document, for example current, draft or superseded, using SharePoint columns or labels, and set new documents to default sensibly.
- We configure the AI index to use only current documents by default, with drafts and archives available only when someone explicitly asks for history.
- Every answer shows the source document's title, date and status next to it, so a reader can see at a glance what they are relying on.
- We set up a regular report of documents with no owner or no status, so the problem does not quietly rebuild itself.
We deliberately do not delete anything. Old versions stay where they are for record-keeping; they just stop being the default source for answers.
What the team notices
Answers come from the document HR or finance actually approved, and staff can see which one it was. When a policy changes, the owner marks the new version current and the assistant picks it up from then on.
As a side effect, people also find things faster without AI, because the clutter is labelled and the current versions are easy to spot.
The ownership report does the long-term work. Each month it lists documents that have lost their owner, usually because someone left, and documents past their review date. A quick look at that list is far cheaper than discovering the gap when the assistant gives a new starter the wrong answer.
Is this your situation?
- Your AI assistant or Copilot has cited drafts or superseded documents.
- The same document exists in several folders with slightly different names.
- Nobody could say quickly which version of a key policy is current.
- Staff double-check AI answers with a colleague before acting on them.
- There is no owner or review date on most of your internal documents.