Confluence AI Assistant
Confluence is where most companies keep what they know and where nobody can find it. Search returns forty pages, three of which are current. An assistant that answers with citations fixes the finding problem — and exposes how much of the wiki is out of date.
Discuss your projectPermissions are the whole problem
Confluence's access model has more layers than most: space permissions, page restrictions that can be set on an individual page, and restrictions that inherit down a page tree. A page about redundancy planning or salary bands sits in a restricted space or carries a page-level restriction, and that restriction is the only thing keeping it private.
An assistant built carelessly loses all of it. Pages are exported, embedded into one index, and the restrictions are gone — so the assistant quotes the restricted page to anyone who asks a question that happens to match it. The content was always protected; the search over it was not.
We carry each page's effective permissions into the index and filter at retrieval against the identity of the person asking. A page the user cannot open in Confluence is not retrieved, so it cannot be quoted, summarised or paraphrased. The model is never shown content the asker is not entitled to — instructing a model to keep a secret it has been told is not access control.
Permissions change, so the sync keeps them current: a page newly restricted or a user removed from a group takes effect on the next sync, and we set that interval against how sensitive the space is.
What it does
Answers with citations
Every statement linked to the page and section it came from, so a user can open the source and check. This is what makes the output trustworthy rather than merely fluent.
Says when it does not know
If retrieval finds nothing relevant it says so and suggests who to ask, rather than filling the gap. The most important design decision for whether people keep using it.
Shows how current the source is
Each answer carries the last-modified date of the pages behind it, so nobody acts on a two-year-old runbook without knowing its age.
Searches across spaces
Within what the user may see, which solves the common case where the answer is in a space the person did not think to look in.
Reads page structure properly
Tables, panels, code blocks and nested lists preserved, and pages chunked by heading so a procedure is not split from its precondition.
Works where your team works
A Confluence macro, a Slack or Teams bot, or both. A separate website to visit gets used once.
Practical notes on Confluence specifically
- Cloud and Data Center both work, with different APIs. We confirm which you run before scoping, since the permission model is read differently.
- Archived and draft pages are excluded by default — answering from a draft is a reliable way to lose user confidence early.
- Attachments are indexed as well as page bodies, because the real content is often a PDF or spreadsheet attached to a thin page.
- Labels and space keys are kept as metadata, so retrieval can be constrained to a space or a label when a user asks about a specific area.
- Jira can be indexed alongside, where the answer to a question lives in a ticket rather than a page.
- Every query and the pages retrieved are logged, which internal audit invariably asks for and which also tells you what people cannot find.
Common questions
Tell us what you need
Describe the problem and we will come back with scope, timeline, and a rate. No obligation to proceed.
Get in touch