An AI agent that fixes documentation errors

Customers report roughly 1,100 documentation errors a month. Each report goes to a writer, who has to reproduce it, find the source file, write the change, and get it reviewed.

By 2026 about 15,000 reports were open. The writing teams were never going to close that by hand, and every open report is a customer who found something wrong.

• An agent that triages each report, writes the documentation change, and opens a code review for the page owner.
• A person approves every change. The agent answers review comments and closes the report after the change merges.
• Start small. The pilot ran one service team's entire backlog, just under 200 reports, before we opened it up to other services.
• Agree early on which kinds of changes the agent may correct by itself and which always need a reviewer.

• The main metric is the share of customer reports the agent fully resolves. Drafted changes don't count toward it.
• I moved the definition of resolved from merged to live on the site. A merged change that customers can't see yet hasn't helped anyone.
• A dashboard tracks volume, approvals, and output, so service teams can see how the agent is doing without asking.

• The agent was code complete and the pilot launched in August 2026. Readiness work for the cross-service rollout followed.
• Next up: send new reports to the agent by default, and show open fixes in the tool owners already use so they can review them there.