AI and Notion.

What Notion AI already does (October 2026)

Notion’s built-in AI covers most “ask the wiki” needs, if you pay for the right plan. On Notion’s pricing page as listed on 19 October 2026 (shown in USD to NZ visitors), Plus costs US$10 per member per month billed yearly and includes only a trial of Notion AI. Business, at US$20 billed yearly or US$24 monthly, adds Notion Agent, AI Meeting Notes and Enterprise Search. Custom Agents, which run repetitive tasks on their own, are free to try and then cost US$10 per 1,000 monthly Notion credits.

A useful decision rule: if staff ask questions inside Notion and everyone is already on Business, start there. Build a custom integration when answers must also draw on other systems, appear in another channel, or write back through an approval step you control.

Integration routes and when to use each

Automation 1: keep the knowledge base current

Stale pages are the main reason AI answers from a wiki go wrong. This job finds them and asks the owner to act; it never edits page bodies.

Inputs

Pages in a “Policies and procedures” data source with Owner, Status and Last reviewed properties, plus newer pages and comments on the same topic found through search.

Model step

For each page not reviewed in 180 days, list what it claims, compare with newer material, and name specific conflicts with links to both sources.

Human approval

The page owner receives a comment listing conflicts and a suggested correction, then accepts, edits or dismisses it.

Write-back

Comment through the API; set Status to “Needs review”. Last reviewed changes only when the owner signs off.

Failure and recovery

Pages the run did not reach keep their old Status, so a filtered view shows what to rerun. Pages moved to Trash through MCP can be restored with notion-restore-pages or from Trash.

Automation 2: permission-aware answers from Notion

The authorization docs set the constraint: an internal connection can read a page only after it is manually shared with that connection, and sharing a parent page shares its children. A Teams or website bot built on one internal connection therefore shows every user everything the connection can see.

1. Pick the access model. For personal assistants, use the hosted MCP server or a public OAuth connection, so each person reads only what they can read. For a shared bot, create one connection per audience (all staff, managers) and share only that audience’s pages with it.
2. Index with metadata. Store page ID, URL, teamspace and last edited time with every chunk, and re-index on webhook events. Our guide to building private AI search covers the indexing design.
3. Cite or decline. Every answer links the pages it used. If retrieval finds nothing relevant, the bot says so rather than guessing.
4. Check before relying on a result. Notion notes that AI search results omit a page’s verification state, so fetch the page itself when that matters.

Permissions and approval design

Give back-office connections read and comment capabilities only, and add write access per workflow rather than by default. On the MCP side, workspace-owned connections expose only the tools an admin selects, MCP tokens can be created only by workspace owners on Business and Enterprise, and limiting which clients members connect needs Enterprise, per Notion’s MCP docs. Do not rely on an MCP client to ask before it writes; put approvals in your own layer, as in our RAG development and AI integration work.

Wiki pages often hold staff and customer personal information. Under the Privacy Act 2020 you remain responsible for it when a provider processes it on your behalf, so record which model providers see Notion content and their retention terms.

Costs and limits (as listed 19 October 2026)

• API: 180 requests per minute per connection, or 600 on Business and Enterprise, plus a shared per-workspace limit. Retry-After is at most 60 seconds.
• Free workspaces: those with more than one member are capped at 1,000 lifetime blocks through the API, effective 8 September 2026.
• Model cost, worked example: a weekly refresh over 400 pages at about 3,000 input and 300 output tokens each, on Claude Sonnet 5.5 at US$2 and US$10 per million tokens, is roughly 400 × US$0.009 = US$3.60 a run.

Where it goes wrong

• Late events. page.content_updated is aggregated and may arrive a minute or two after an edit. Add a nightly pass on last edited time to catch anything missed.
• Partial pages. Very large pages can come back from notion-fetch with truncated set. Check the flag before letting the model answer from them.
• Rate-limit errors. Notion MCP retries once when the wait is two seconds or less; otherwise it returns a rate_limited tool error. Agents that fan out in parallel hit this first.
• Instructions inside pages. Pages edited by guests or pasted from email can contain text aimed at the model. Treat page content as data; see recognising prompt injection.

Teams holding documents in more than one place should read our SharePoint and Google Drive guides, since cross-system search changes the permission design.

Written by the Levy Street team in Wellington. Last updated 19 October 2026.