Knowledge Management AI: Rovo, Notion AI and Guru on Permissions and Price

Atlassian gives each Confluence Standard user 25 Rovo credits a month, and one basic chat message costs 10 of them. That is 2.5 messages per user before the meter starts. Knowledge management AI now ships inside every major wiki, and price turns out to be the smaller decision. The decision that matters is whether the assistant shows each person only what they could already open.

What knowledge management AI is actually selling you

Every product in this comparison does the same visible thing. You type a question, and an assistant answers it from your company's documents with citations. The demo is close to identical across all three vendors.

What differs sits underneath. Each assistant needs a copy of your content, or live access to it, plus a record of who is allowed to see each document. The answer is only as safe as that record is current.

That makes document-level access the deciding factor, not answer quality. A wiki assistant that summarises a restricted salary review page for the whole company has failed regardless of how fluent the summary was. The same problem in the enterprise search category is covered in the comparison of Glean and Microsoft Copilot on permission plumbing. This post applies that lens to the three wiki and knowledge base vendors most teams already pay for.

The three are not identical products. Confluence with Rovo is a wiki with an assistant bolted onto Atlassian's index. Notion AI is a workspace with an assistant that can also reach outside it. Guru sells itself as a governed knowledge layer that sits across your other systems, Confluence included.

Our position, stated up front: if your company already runs on Confluence, Rovo is the default and the burden of proof sits with the alternatives. The rest of this post is about when that burden is met.

Rovo, Notion AI and Guru pricing, as each vendor publishes it

Pricing transparency varies more than the products do. One vendor publishes a seat price and a meter. One publishes a meter and tiered seat pricing that changes with headcount. One publishes neither.

Confluence per-seat dollar figures are deliberately absent. Atlassian's pricing page renders them in a calculator that varies by user count, and third-party sites quoting a single figure disagree with each other. Read the current price for your headcount from Atlassian directly.

Confluence: the seat is cheap, the AI is metered

Atlassian made Rovo part of every paid cloud plan, with no separate purchase. The cost appears instead as a credit allowance. Atlassian's own credit documentation lists 25 credits per user per month on Standard, 70 on Premium and 150 on Enterprise.

The seat price itself moved recently. Atlassian's investor notice on fiscal 2026 cloud pricing listed increases of 5% on Confluence Standard, 7.5% on Premium and 7.5% to 10% on Enterprise, effective 15 October 2025. Those increases phase in at renewal, so many customers are meeting them now.

Notion: one published number, plus a second meter

Notion folded its AI into the Business plan. The pricing page lists Business at $20 per member per month, with a billing toggle noting that yearly payment saves up to 20%. Notion Agent, Enterprise Search and AI connectors sit at that tier.

A second meter exists for automation. Custom Agents are "free to try, then $10 per 1,000 monthly Notion credits", and Workers carry the same requirement. That unit price matches Rovo's overage rate exactly, which is either coincidence or a sign of where the market is settling. The broader confusion that credit pricing creates for buyers is laid out in the breakdown of credits, tokens and AI billing units.

Guru: priced as a project, not a seat

Guru's pricing page carries no plans and no figures. It describes Guru as "a platform and expertise solution, not just a per-seat tool", with pricing "tailored to your organization's scale, knowledge complexity, and AI maturity".

Third-party sites still quote an old self-serve seat price. We did not use it, because Guru's own page no longer supports it. A buyer comparing the three on a spreadsheet therefore has two real numbers and one blank, and the blank is the vendor's choice.

The Rovo credit meter in practice

A pooled allowance is easy to misread, so it helps to convert credits into actions. Atlassian charges 10 credits for each basic event, such as a Quick Answers message in Rovo Chat or a basic Rovo Agent. "Think deeper" chat and advanced agents carry variable pricing per event.

Several things stay free after the allowance runs out. Atlassian lists Rovo Search inside Atlassian apps, page summaries and definitions as free. Storing indexed objects from connectors does not use credits either.

Now run the arithmetic for a realistic team. A 100-seat Confluence Standard site gets 2,500 pooled credits a month, which covers 250 basic chat messages for the whole company. If each person sends 20 basic messages a month, usage is 20,000 credits. The 17,500 credits above the allowance cost $175 at $0.01 each, or $1.75 per user. Those usage figures are illustrative, and variable-priced events would raise the total.

Our view is that the overage cost is the least important number in this comparison. At list price, heavy basic usage adds a modest sum per seat. The meter matters more as a governance signal: extra usage is on by default, and admins receive alerts at 80% and 100% of the allowance. Whether you prefer this model to an all-inclusive seat is the argument in the piece on bundled versus metered AI pricing.

The quieter risk is behavioural. A small, visible allowance teaches people to ration questions, and rationed assistants rarely become habits. If you buy Rovo for adoption, budget the overage deliberately rather than hoping the pool lasts. Vendors tightening allowances after launch is a pattern documented in the analysis of AI usage limit shrinkflation.

How each knowledge management AI handles document-level access

Document-level access means the assistant checks, for each document it reads, whether the person asking could open that document in the source tool. All three vendors say they do this for their own content. The differences appear when content comes from somewhere else.

Rovo: inherits by default, publishes no sync clock

For Atlassian content, the rule is simple. Atlassian's overview says "Rovo respects your user permissions, so users can only see what they already have access to." Confluence page restrictions carry straight through.

Third-party content works through what Atlassian now calls Teamwork Graph connectors, of three types. Synced connectors "store and index all the content in a workspace using a service or admin account", alongside the permission settings for each file. Direct connectors query the source live. Smart Link connectors only surface links a user has already viewed.

Atlassian's permission sync page says synced content can only be seen by users who already have access in the source app. It gives no interval. The data privacy page says Atlassian "listens" for permission changes, and uses softer wording elsewhere: syncing "helps ensure" users only see content they can access.

Two connector details matter in practice. Google Drive results require each user to connect their own Google account first. Slack connections made after 2 October 2025 are Direct, so Atlassian states "your Slack messages are not indexed by Atlassian", covering public and private channels and direct messages. Older Slack connections still index content, and switching them over is documented as irreversible.

Notion AI: inherits, with documented timings

Notion's language is the most direct of the three. Its connector help page states that "users will not be able to generate content or receive responses based on resources they do not have access to." For actions, the agent "acts as you, so it can only access what you can already see."

Notion also publishes the timings the others omit. Initial ingest can take up to 36 hours. New content "may take up to 3 hours" to appear. The Google Drive connector syncs new content "about once an hour", and Slack edits and deletions arrive by webhook "in near real-time".

The edge cases are documented too, which is to Notion's credit. In Google Drive, files shared with "Anyone with link" must be opened at least once before Notion AI can find them, and target audiences are not supported. In Slack, an admin selects public channels while each user connects their own private channels and messages. Slack Connect content is excluded from the index, yet real-time search can still surface Slack Connect channels you belong to.

One stance worth taking: Notion's one-year lookback on connected apps is a sensible default, not a weakness. Old Slack threads are where stale and wrong answers come from, a failure mode examined in the post on stale data behind agent answers.

Guru: two models, and the admin picks

Guru's own content uses roles on Collections, folders and Cards, with Viewer, Owner, Creator and Admin roles plus custom roles. That part is conventional.

Connected sources are where Guru departs from the other two. Its source documentation describes two models. In the Guru Groups model, "Guru connects as a single user and indexes only what that user can access. Access is controlled entirely within Guru, independent of Source app permissions." In the inherited model, "users will only see content in Guru that they already have permission to access in the Source app", and that model "is supported for some sources."

The per-source pages show the cost of inheriting. Per-user Google Drive visibility needs Domain-Wide Delegation, which requires a Google Super Admin. The SharePoint page states that because of a Microsoft Graph limitation, "SharePoint Group memberships cannot be retrieved", so permissions will not sync automatically for group-only sites. Guru's Confluence source syncs every two hours.

There is also a gap between Guru's marketing and its manual. The pricing page says permissions are "inherited from your existing systems and enforced in real time." The help centre says Guru "checks for changes multiple times per day." We would trust the help centre, and we would not sign a Guru contract that puts a sensitive source on the Groups model without a written reason. Where an assistant connects through a shared account rather than each user's identity, the risks follow the pattern in the analysis of agent credentials versus user impersonation.

Connector lists, and what each connector actually reads

Connector counts are the easiest number to compare and the least useful. Each vendor counts differently, and two of them contradict their own pages.

Atlassian's setup documentation carries 32 "Connect to Teamwork Graph" guides, plus 31 apps searchable through Smart Links. Notion's connector help page lists 10 apps, while its pricing page names Box, Salesforce and Asana as beta connectors that the help page omits. Guru markets "100+ integrations", and its own source list in the help centre runs to 72 entries, several of them variants of the same product.

"Listed" means the vendor documents the connector. It does not confirm that the connector inherits source permissions. Guru does not say which sources support its inherited model in the source list itself.

The table's last row matters as much as the others. Guru is the only one of the three that indexes Confluence as a source, so a company can run Confluence and Guru together. Rovo needs no connector for Atlassian content at all. Notion publishes latency figures nobody else does. Each wins somewhere, and a company with heavy app sprawl should weigh the source list hardest, a problem sized in the analysis of companies running 291 apps.

Where this comparison is weakest

This post reads documentation. That is a real limitation, and the reader deserves it stated plainly rather than buried.

Documentation is not behaviour

Every permission claim above is what a vendor wrote, not what we measured. Vendors document intended behaviour, and the interesting failures happen outside it. Notion's two pages already disagree about which connectors exist, and Guru's marketing and manual disagree about timing. A document that contradicts itself on one point may be stale on others.

The price comparison is also incomplete. We could not read Atlassian's per-seat dollar prices from an official page, because they render inside a calculator, and Guru publishes no price at all. A buyer's real quote for any of the three will differ from list, often substantially at volume.

The case for Guru's decoupled model

The strongest objection to this post's framing is that inheriting source permissions is not always the goal. Many support and sales teams want a curated answer set that deliberately ignores the mess in shared drives. For them, an index controlled inside Guru is the product, not a gap.

That view genuinely wins in one situation: when the source systems are overshared and nobody will fix them before launch. Inheriting broken permissions faithfully reproduces the breakage at the speed of a chat answer. A decoupled index, curated by an owner, can be safer than a faithful copy of a bad access model. We still think inherited permissions should be the default, because curation depends on one admin staying careful forever.

A permission test to run before you sign

Documentation tells you what to test. Only a test tells you what happens. Run the following in a trial workspace with two test accounts, one with access and one without, before any contract is signed.

Write down the revocation lag you observe for each vendor. That single number is more useful than any connector count, and none of the three vendors publishes it for permission changes. If a vendor refuses a trial long enough to run these six tests, treat that as the answer.

Frequently asked questions

Where to start

One test and one number, both small enough for this week.

Run the restricted page test from the table above in whichever assistant your company already has switched on. Most teams have never checked, and Rovo is on by default for every paid Atlassian cloud plan.

Then open your admin console and find the AI usage or credit view. If you run Confluence, compare last month's consumption against 25, 70 or 150 credits times your seat count. That comparison tells you whether 3 December 2026 is a non-event or a budget line.

Related reading