Every editorial site has a drafts folder that only grows. We audited ours on 19 November 2026 and the result was blunt: every unpublished article on the site, 7 out of 7, had gone untouched for more than 90 days. None of them was ready to publish as it stood.
What we checked
The rule was simple. An article counted as stale if it was unpublished and its last change was before 21 August 2026, 90 days before the audit. For each draft we recorded the title, author and last-changed date, then read the draft itself and decided whether to finish, publish or delete it. We based the verdict only on what the draft says about its own state, not on guesses about what the editors intended.
The findings
| Draft | Last changed | Verdict |
|---|---|---|
| Draft: festival coverage plan | 21 Jan 2026 | Delete |
| Draft: summer preview | 25 Mar 2026 | Delete |
| Utkast: sammenslåing av to studieprogram i humaniora | 23 Mar 2026 | Delete (low confidence) |
| This is the first AI article | 27 Jul 2026 | Delete, unless it is a demo fixture |
| Draft: interview with a first-time director | 22 Apr 2026 | Finish |
| Utkast: nye retningslinjer for praksisavtaler | 20 Apr 2026 | Finish |
| Utkast: pressemelding om nytt forskningssenter | 19 Jan 2026 | Finish only if funding was confirmed |
Four to delete, three to finish, zero to publish.
Why they stalled
The drafts fall into three groups.
- Overtaken by events. The festival plan was replaced by a later decision and the draft itself says to delete it. The summer preview had dates that moved before it was finished, and summer is now over. The study programme merger was postponed by the faculty board, and the text was never updated after that.
- Blocked on someone else. The press release cannot go out before the grant letter arrives, and the rector's quotes are not approved. The practice agreement guidelines are waiting for review by the legal section. The director interview needs a confirmed release date, a cleared photo and the interviewee's consent to one quote.
- Test content. "This is the first AI article" is generic generated text with a placeholder title. It was most likely left over from early AI feature testing.
The pattern is clear. Nobody abandoned these drafts on purpose. Each one was waiting for an outside trigger: a decision, an approval, a letter. When that trigger never came, or came and nobody went back to the draft, the draft just stayed where it was.
How the audit ran
The audit was done by an AI agent connected to the site through its MCP tools. The connection allows edits, but for the audit the agent was told to change nothing, and it only read content. It listed unpublished articles sorted by last change, applied the 90-day cutoff, read each body and drafted the verdicts. A person still has to make the final call. The agent can tell you that a draft says "do not publish before the grant letter arrives", but it cannot tell you whether the letter ever arrived.
What we will do differently
- Name the blocker in the draft. The drafts that were easiest to judge were the ones that said what they were waiting for. That should be standard practice.
- Put a review date on anything blocked. A draft that waits for approval needs a date when someone checks on it again.
- Run the audit every quarter. Reading seven drafts takes minutes. Reading seventy after two years of neglect does not.
- Delete without guilt. A draft that events have overtaken has no value left. Revision history keeps the text if anyone ever needs it.