Can I control which articles the AI uses?
Yes — three places, in decreasing order of bluntness.
1. Article status — the kill switch
Only articles with status published are retrievable by any AI agent.
draft→ invisible to retrieval. Use this for new articles you’re still writing or AI-generated drafts you haven’t approved.needs-review→ invisible to retrieval. Use this for published articles you’ve spotted as wrong; the AI stops using them until you re-approve.archived→ invisible to retrieval. Use this for articles whose feature is gone.published→ live in retrieval.
Changing status is the fastest, most surgical control. An article you want pulled from the AI immediately: set to needs-review, save. Within a minute the retrieval engine drops it.
2. Category access — hide a category from agents or grant it to chosen ones
On a Knowledge category, the shield icon opens an Agent Stack access panel. It controls which Agent Stacks can retrieve from that category.
- Confidential — The category is hidden from every Agent Stack. No specialist retrieves its articles, whatever its own knowledge scope says.
- Stacks or specialists ticked — Adds access for the ticked stacks or specialists that read only chosen categories. Specialists that read the whole knowledge base still read the category.
Use Confidential for content Agent Stacks should never draw on, such as internal-only policies; it does not limit Sidekick or the public site. Use ticking when a specialist is scoped to chosen categories and needs one more category: ticking its stack leaves the specialist’s own scope as it is, while ticking the specialist adds the category to that scope.
Confidential applies to the whole category, not to individual articles.
3. Specialist knowledge scope — narrow what each agent can see
When you configure a specialist (the agent inside an Agent Stack), you can scope its KB access on its Knowledge tab:
- Whole knowledge base — The specialist retrieves from your entire published KB, except categories marked Confidential.
- Chosen categories, whole categories checked — The specialist retrieves articles in each checked category, including articles in its current and future subcategories. Articles outside those category trees are invisible to this agent, even though they’re still public on the help center, unless they’re in the Supervisor’s Self-learning category or a category ticked for its Agent Stack.
- Chosen categories, single subcategories checked — The specialist retrieves only articles assigned to one of the checked subcategories, plus the Supervisor’s Self-learning category and any category ticked for its Agent Stack. Checking a subcategory does not include sibling subcategories or the whole parent category.
This is the right lever when you want different agents to specialize. A billing agent can get the whole Billing category, including any Billing subcategories you add later. Or, if the agent should answer only refund questions, you can check just the Refunds subcategory so it retrieves only articles assigned there. The KB stays whole, but each agent retrieves only from its scoped slice.
Scope is set on a specialist’s Knowledge tab when you build or edit it. See the Capabilities docs for the configuration walkthrough.
The KB editor and the Supervisor publish destination picker also help catch unreachable content. If no Agent Stack specialist can retrieve an article at the selected category and subcategory, Atender warns you before save or publish.
Sections — group articles for the public help center, not for AI
Sections are an alternative grouping for the public help center — a way to surface “Top articles” or “Getting started” boxes on the home page. They don’t change retrieval. The AI doesn’t know about sections; it sees the same article whether or not it belongs to a section.
If you wanted sections to affect retrieval, the right answer is specialist knowledge scope on a category that maps to the same content.
What does not control AI retrieval
Two things that sound like they should but don’t:
- Roles — KB roles are browsing-visibility tags for the public help center. They control who sees what when browsing. The retrieval engine ignores them — it treats every published article as a candidate regardless of role tags.
- Tags — Tags filter articles in the in-app editor and on the public help center. They don’t change retrieval scores.
These can feel like missing levers. The framing fix is: status, category access, and specialist knowledge scope are the levers, and they’re enough. If you find yourself wanting role-bias on retrieval, the better answer is usually a specialist scoped to the right category.
Edge cases worth knowing
- A drafted article that was previously published. When you flip status from
publishedtodraft, the article drops from retrieval within a minute. In-flight conversations finish with whatever they already retrieved. - A category renamed mid-conversation. Doesn’t affect retrieval; rename freely.
- An article moved between categories. The article stays retrievable; its category scope changes for any specialist set to “Chosen categories”. If you move it into or out of a Confidential category, its availability to agents changes with the move.
- A Confidential category and a specialist that reads the whole knowledge base. The category stays hidden. Confidential hides it from every Agent Stack, including whole-KB specialists.
- Unpublishing and republishing. Doesn’t trigger a fresh embedding pass; the article’s existing embeddings are reused, because retrieval filters on status at search time.
What about per-article overrides for specific agents?
There’s no per-article “exclude from agent X” toggle. Category-level Confidential is available, but it works on a whole category, not on a single article. The three levers above — status, category access, and specialist knowledge scope — are the design. If you need finer-grained control than “scope this specialist to category Y” or “mark this category Confidential”, consider:
- Splitting the article into two so each part lives in a category that maps cleanly to its agent, or into a Confidential category if no agent should use it.
- Using the Handbook for internal procedures the AI follows but never quotes. Handbook entries are agent-readable but never customer-facing.