An unanswered question is not automatically a missing-document problem. The information may exist but be inaccessible, poorly indexed, ambiguous, or split across sources the assistant failed to connect.
Before creating documentation tasks, diagnose why the assistant could not answer. Otherwise, a backlog fills with duplicate articles while the actual problem remains in retrieval or permissions.
Triage the evidence gap before assigning writing work
Capture the question, intended reader, authorised source set, retrieved passages, and refusal or incomplete answer. Classify the result as a genuine source gap, retrieval miss, permission boundary, unclear question, conflicting documentation, or unsupported requested action.
Use a human source owner to confirm genuine gaps. A search result that contains no answer proves only that this search found none. It does not establish that the whole organisation lacks the information.
Trace a fictional onboarding question into a task
Imagine a new employee asks where to send a replacement-laptop request. The assistant returns a general IT policy with no request channel. A reviewer finds an old team message mentioning an address, but no maintained policy confirms it.
The documentation task should ask the IT owner to establish the current request channel, eligible users, required information, and escalation route. Include the employee's underlying task in plain language. Do not ask someone merely to “add content for AI,” which gives the owner no acceptance criterion.
The task is complete when an authorised employee can find a reviewed source that answers the request-channel question and the source has an owner and update date. Add a regression question that uses different wording from the new page heading. That checks discoverability rather than memorisation of a title.
If the maintained channel already existed but retrieval missed it, create an indexing or search task instead. If it is restricted for a reason, improve the assistant's explanation of the permitted next step without copying protected content into a public page.
Prioritise by recurring task and consequence
Group similar questions while preserving meaningful differences in department, geography, or product. “How do I request access?” may describe several distinct approval workflows. Merging them too aggressively can produce one broad page that answers none of them well.
Prioritise gaps using recurrence, blocked work, owner confidence, and risk of a wrong answer. Raw question counts can overrepresent a temporary incident or a few enthusiastic users. Review representative examples before deciding what to write.
This workflow also has a privacy boundary. Remove personal details from documentation tickets and restrict the underlying evaluation trace. The goal is to improve reusable instructions, not distribute private employee enquiries.
See connecting company documents for integration context and the Nucleus project for related portfolio work.
Questions about documentation backlogs
Should every refusal create a ticket?
No. Some refusals are correct, and some questions are outside scope. Triage and group them before assigning work to source owners.
Who decides the new answer is authoritative?
The business owner of the policy or workflow. An engineer can improve retrieval but should not invent operational rules to fill a gap.
How do we close the loop?
Link the published source revision to the original case, rerun the question as the intended reader, and check the next action. Use unresolved cases in a technical review.