Quick answer
Most knowledge bases go stale for a simple reason: nobody knows which articles to write next, so the gaps only surface as tickets. Closed-loop knowledge mining inverts that. Resolved conversations are the best record of what customers actually ask, and Orbit already ranks those reasons on the automation-opportunities surface. Closed-loop mining takes the recurring reasons flagged as best answered by a help article, drafts the article from the real customer wording, and files it into a review queue — nothing reaches your AI agents until a human approves it. The loop is closed: conversations teach the knowledge base, and the knowledge base answers the next conversation.
The opt-in and controls live at `/insights/kb-auto-draft`: a toggle for the 24h scheduled sweep, a per-run draft cap, and a Run now button that triggers a mining pass on demand even while the sweep is off.
Why a closed loop, not a writing sprint
The typical knowledge-base program is a quarterly sprint: a team blocks a day, writes articles from memory, and publishes them. Two problems follow. The topics come from what the team remembers, not what customers asked last month. And the article backlog competes with every other priority, so the KB drifts stale again within a quarter.
Mining resolved conversations fixes the topic problem first. The automation-opportunities surface already reads every resolved conversation and ranks the recurring contact reasons that a help article would deflect well. The auto-drafter works from that ranking, so the articles that appear in your review queue are the ones carrying the most volume — the ones a writing sprint would have guessed at.
The draft only shifts the editorial cost; it does not remove it. Deterministic, human-reviewed drafting means the reviewer's job is to write the resolution, not to repair a hallucinated one.
What the dashboard surface controls
Open /insights/kb-auto-draft and there are two cards.
Auto-draft settings. A toggle for the scheduled sweep plus a Max drafts per run field — between 1 and 10, defaulting to 3. The cap keeps one sweep from flooding the review queue; three drafts an editor can actually read beats ten they skim.
Run the miner now. A Run now button that triggers a mining pass for your organization on demand. The manual run works even while the scheduled sweep is off — useful for a one-off harvest before you decide whether to leave the sweep on. The page links straight into your knowledge-base page, where the drafts wait in the moderation queue.
Owners, admins, and developers see the page; the run action is a write, so it is restricted to those roles.
What a draft looks like when it lands
Each mined topic produces one draft article, filed into a dedicated Auto-Mined Drafts (Pending Review) knowledge base that Orbit creates on the first run. The draft is assembled extractively, not generated — no model rewrites your customers' words:
- Why customers are asking. The conversation volume for this reason over the mined window, the channel it mostly arrives on, and the current self-service deflection rate for it.
- What customers are saying. Up to five deduplicated, truncated excerpts from real resolved conversations, so the phrasing customers actually use survives into the reviewer's draft.
- A reviewer checklist. Three unchecked items: replace the excerpts with a clean explanation, add the real resolution steps, then approve to publish and index for your AI agents.
The title is marked as auto-mined in the document name, and the topic it sprang from rides along on the document, which is what keeps the miner from re-drafting a reason you already have in the queue — approved, pending, or rejected.
Determinism is the point: run the same conversations twice and you get the same draft. The reviewer is editing real excerpts into an article, not fact-checking a probabilistic summary.
Review, approve, publish: the moderation path
Nothing in the Auto-Mined Drafts pool is visible to AI agents on arrival. The pool is created with a required publish-approval step, so every draft — however it arrived — sits at pending until someone opens it in the knowledge-base moderation queue and approves or rejects it.
Approval is the moment the draft becomes useful: on approval the article's content is chunked, embedded, and indexed, and from then on your AI agents can retrieve it to ground an answer. Rejection keeps the topic's block list clean — a rejected draft still counts as "already drafted," so the sweep will not re-file the same reason next run.
That is the whole moderation path: mine → draft → review → approve → indexed. Two of the five steps are human.
Numbers worth watching afterwards
An approved draft is a hypothesis, and Orbit gives you two readings on whether it held.
The deflection rate moves. The drafts quote the pre-draft deflection rate for their topic; the automation-opportunities ranking keeps scoring it after publication. Drop a draft's source reason on a second look and it should sit noticeably higher.
The queue stays small. If Run now keeps returning zero new drafts, the miner has run out of uncovered high-volume reasons — a good sign, not a broken one. The sweep is quiet when the knowledge base is caught up.
If a topic recurs despite an article, read the excerpts again — the article answered a different question than the one customers arrived with.
Turn it on
Open [`/insights/kb-auto-draft`](https://orbit.devotel.io) in the dashboard, run the miner once, and read what it files. If the drafts are worth an editor's afternoon, flip the sweep on and leave the cap at 3; if they are not, the toggle is the whole commitment.