AI Agent Connected to Confluence Knowledge Base

An AI agent connected to a Confluence knowledge base reads what its connection is scoped to, and that scope is a product setting, not a constant. Atlassian states that a Rovo agent has access to all organizational knowledge unless it is narrowed to chosen spaces or a branch of the content tree. Moveworks offers selected spaces, all spaces except some, or all spaces, and recommends the first. Serval and Harmony have an administrator pick spaces and pages. Inside that scope, Atlassian, Serval and Moveworks state that Confluence permissions also filter the answer, so a user only receives content they may view in the source. Freshness depends on the connector: Serval states a sync every four hours, Ravenna every twenty-four, and Moveworks describes real-time sync after a first crawl of a few hours to 48. Serval and Moveworks state that their sync is read-only. Four questions decide the rest: which spaces, whose permissions, how fresh, and what happens when nothing answers.

What does a Confluence connection actually take in?

Scope is set at two levels: which spaces the connector ingests, and which content inside them it keeps.

• Atlassian Rovo. A Rovo agent “has access to all organizational knowledge unless you configure it differently”. Choosing custom knowledge narrows it to Confluence spaces, or to a branch of the content tree, where selecting a parent adds “all its child content”. A third setting, no organizational knowledge, removes it.
• Moveworks. Space selection is mandatory, with three options: only selected spaces (recommended), all except selected, or all spaces. Pages and blog posts with no body content are excluded, so a title-only page used as a section container never appears in results.
• Serval syncs Confluence “spaces and pages”, chosen item by item, and advises excluding drafts, internal notes and outdated content from what is synced.
• Harmony opens a tree view of spaces and pages, from which an administrator selects items and starts the sync.
• Ravenna keeps the hierarchy of the source, “Confluence space structures” included, inside knowledge folders that are themselves scoped: an agent searches only the folders connected to it.
• Siit’s Confluence page has an administrator pick the spaces to import and converts each page into an article with its own state. Drafts are excluded from the AI, so a synced page is used only once it is published in Siit. The same page warns that some Confluence macros, “especially third-party ones”, may not render, which leaves a synced article empty or partial.

“Connected to Confluence” therefore says little on its own: two products connected to the same site can read different sets of pages.

Whose permissions decide what an answer can use?

Confluence’s own model comes first. Atlassian describes its permissions as “open by default”: content is available to everyone with view access to the space unless someone restricts it. A page restriction passes downwards for viewing: “If someone can’t view a parent content item or folder, they won’t be able to view any child content items under it.” Editing restrictions do not work this way. Atlassian also states that content cannot be restricted on the free plan.

On top of that model, the vendors read here describe two filters.

The reach of the connecting identity. Moveworks ingests only what its service account can read; with its Forge app, public spaces are ingested by default and restricted content only once the app is granted access. Siit writes that pages the connected user can’t access won’t appear, and recommends a dedicated service user because Atlassian access is tied to a user. Serval shows a “cannot be reached” warning naming the Confluence spaces an administrator has restricted, and that warning does not clear until access is restored. A restriction placed high in a space can therefore take a whole branch out of what the agent ever sees.

The person asking. Rovo “respects the permissions of the user who prompted it because it is acting on their behalf”. Serval states that the agent “only surfaces a document to a user who is allowed to see it in the source system”, and re-reads synced permissions when content changes. Moveworks states that restricted spaces and pages remain restricted in search results. Harmony takes permissions from the source, Confluence space permissions included, shows the authorized groups per article, and flags “Some groups are not synced” when a group from the source is missing from its directory.

Atlassian’s own service desk carries an exception worth knowing. In the Jira Service Management virtual service agent, AI answers reflect the customer’s permissions on knowledge base articles, “except in Slack, where only articles set to All logged-in users will be used for AI answers”. Activating AI answers also requires the linked knowledge base space to be set to All logged-in users.

What changes when the documentation lives in Notion or Slab?

Notion. Notion attaches capabilities to each connection: “Read content”, “Update content” and “Insert content” are separate, and its advice is to “request minimum capabilities that your connection needs in order to function”. A connection that reads content to export it out of Notion only needs read content. A second rule matters as much: when a connection receives access to a page or database, it reaches “both that resource and its children”. What the agent reads is the list it was added to, plus what sits underneath. Serval syncs Notion “pages and databases”, Ravenna preserves “Notion page trees”, and Harmony lists Notion among its third-party sources.

Slab. Slab states that the content its GraphQL API returns “depends on what the Slab Bot user has access to”, and advises checking that account first when content is missing. Slab offers its API and webhooks on the Business and Enterprise plans. For any agent that reads Slab through that API, the Slab Bot’s access is the ceiling and the plan is a precondition.

Serval and Ravenna also list Google Drive next to Confluence and Notion; the way Drive sharing reaches an agent is covered in Google Drive as a knowledge source.

How fresh are the answers after an edit?

What reaches an answer depends on how each connector syncs, and the vendors publish different things about it.

Agent	What it publishes on refresh
Harmony	Articles “sync periodically”; status per article (Completed, Pending, Failed), with the failure reason on hover
Moveworks	Content kept “in sync in real time, ensuring that updates and permission changes are immediately reflected in search results”; the first crawl takes “anywhere from few hours to 48 hours” depending on data size
Ravenna	Auto-sync every 24 hours by default, at a randomised time within the period; manual sync available; failed syncs retry at the next run
Serval	Connected sources sync every four hours; manual sync available
Siit	Content “refreshes on a schedule”, with no interval; a manual Synchronize button, hidden while a sync runs and for 3 hours after the last one

The Atlassian pages read for this guide, Rovo’s included, state no refresh interval, and neither do the Notion and Slab pages read: where an interval is stated, it comes from the connector’s vendor.

After a successful sync, a corrected page becomes available for retrieval if it is in scope, visible and permitted. Moveworks states that updates are immediately reflected in search results. Siit’s page states that on the next sync every Siit surface reflects the update, with no manual republishing. Serval offers a manual sync “instead of waiting for the next scheduled run”, Siit’s Synchronize button is for a page that must be “reflected in Siit straight away”, and Ravenna states that manual sync is available at any time.

What does the agent do when nothing in the wiki answers?

The vendors read do not describe it the same way, and two document a setting.

Harmony exposes a toggle, “Enhance with general AI knowledge”. In knowledge-base-only mode, if no relevant answer is found, the agent escalates the ticket to a human agent rather than answering from general training data; the default mode uses the knowledge base first and supplements it with general trained knowledge. Rovo lets an agent be configured with no organizational knowledge, and its web search can be turned on as an alternative or in addition to the defined scope. Ravenna writes that without knowledge the agent “relies solely on its base training and rules”.

Atlassian’s virtual service agent first tries to match an intent, and only then tries AI answers over the linked knowledge base spaces. Atlassian writes that AI answers work best for questions that can be resolved with information or instructions and do not usually need to be escalated to an agent. Serval advises testing before go-live with the questions users would ask, confirming the agent returns accurate answers with links to the source document.

The practical question is what each product does when its sources run out, and whether that behavior can be set.

Where the connection stops

Deployment type is not settled the same way everywhere. Moveworks documents a separate Confluence Server and Data Center path, where permissions-aware results require installing the Moveworks app in Confluence, and its newer enterprise search configuration is available only for Confluence Cloud. Atlassian states that the virtual service agent and AI answers are not available in the Atlassian Government environment.

Other limits are about reach and credentials. Serval reports “Some content isn’t searchable yet” when a sync could not reach part of a source, often because of a rate limit or missing access. Ravenna keeps a document that disappears from its source, marked with an error, and states that sync fails until expired credentials are reconnected. Siit lists a “Connection lost” state when the Atlassian token is revoked or expired.

The content gate holds as well: Ravenna lists outdated or contradictory documents among the causes of poor answers. Where documentation lives in a ticketing tool rather than a wiki, the handover is covered in an AI agent in front of Jira Service Management, and the work of preparing documents is in training an agent on internal documentation.

• 01

  Scope defaults differ

  Atlassian states that a Rovo agent has access to all organizational knowledge unless configured otherwise. Moveworks makes space selection mandatory and recommends a list of selected spaces.

• 02

  Restrictions travel downwards

  Atlassian states that someone who cannot view a parent item cannot view any child item under it. One restriction near the top of a space can take a whole branch out of what a connected account reads.

• 03

  Read access is a capability

  Notion holds read, update and insert content as separate capabilities of a connection, and advises requesting the minimum needed. A connection that only reads can be reviewed as exactly that.

• 04

  Refresh behavior differs

  Serval states four hours, Ravenna twenty-four, Harmony says periodically and Siit gives a schedule without an interval. Moveworks states that updates are immediately reflected in search results.

Which vendors document how their agent reads Confluence?

Listed here: vendors whose public documentation, read for this guide, states at least two of the three things a Confluence connection decides, namely which spaces are read, how Confluence permissions apply, and how often content refreshes.

• Atlassian: Atlassian’s Rovo knowledge page states that subagents can be scoped to Confluence spaces or a content-tree branch, and that answers respect the permissions of the user who prompted the agent.
• Freshworks: Freshservice’s connect-your-apps article describes choosing the Confluence spaces the AI agent learns from, and says the agent “only surfaces content that the asking user is allowed to see in the original app”.
• Harmony: Harmony’s knowledge base page describes selecting Confluence spaces and pages, permissions taken from Confluence space permissions, and periodic sync with a status per article.
• Ravenna: Ravenna’s Confluence page imports a space by its space key, and its knowledge overview says knowledge folders “auto-sync every 24 hours by default”.
• Serval: Serval’s connected-sources page describes syncing Confluence spaces and pages read-only, per-user permission filtering, and a four-hour sync.
• Siit: Siit’s Confluence page describes picking the spaces to import, honoring Confluence space and page restrictions, and a scheduled refresh with a manual Synchronize option.