Notion AI

The question with Notion AI is no longer "which AI assistant should we bolt onto our workspace?" Since May 2025, Notion has folded its AI features into the Business and Enterprise plans and stopped selling the standalone $10 per-member add-on to new customers. At time of writing, the pricing page lists Business at $20 per member per month billed yearly, with AI included; third-party pricing checks put the monthly-billed figure at $24. Free and Plus workspaces get a limited trial and nothing more. That single change reframes the entire category entry for this product: Notion AI is not a tool you buy, it is a reason to upgrade everyone in your workspace to a plan that costs double the one below it. I have spent the past weeks treating it as exactly that — an upgrade decision, not a tool purchase — and this review is written for the person signing that invoice.

Positioning: an AI layer, not a workspace

First, the distinction I keep seeing blurred in catalog listings: Notion AI is not the Notion workspace. The workspace — documents, databases, wikis, Notion Calendar, Notion Mail — is the base product millions of people already use. Notion AI is the layer sitting on top of it: a chat agent that answers questions grounded in your workspace, inline writing and translation tools, database autofill, AI Meeting Notes, and — on Business and up — Enterprise Search and Research mode. What I look for in this category is whether an AI feature can reach the company's actual knowledge or whether it is a text generator with amnesia. Notion AI's whole bet is that the knowledge is already in the building. When it works, it works for that reason, and when it fails, it fails for the inverse reason — your knowledge was never in Notion to begin with.

The company behind it is Notion Labs, founded in San Francisco in 2013 by Ivan Zhao and Simon Last, among others. Notion has reported more than 100 million users worldwide, a figure repeated in the Cerebras partnership announcement of July 2025 (I treat vendor-reported user counts as claims, not audited facts), and a January 2026 tender offer valued the company at $11 billion. This matters for a review in a way that vendor fluff never does: Notion AI runs on frontier models the company swaps underneath you — the Notion 3.0 release notes name Claude Sonnet 4 and GPT-5 as built-in options — and the survival odds of the wrapper are much better than the survival odds of a three-person AI startup reselling the same models.

What it actually does

Notion 3.0, released on 18 September 2025, rebuilt Notion AI around what Notion calls Agents. The personal Agent is the chat surface: it answers questions with citations from your pages and databases, drafts and edits content, works with database rows, and translates. Notion claims multi-step runs of over 20 minutes with memory stored in a profile page you can edit — treat "20 minutes" as a vendor ceiling, not a guarantee. You write the Agent an instructions page describing your style and filing conventions and it applies them over time. In the same release Notion added AI formula writing: describe the property logic you want in plain language and it writes the formula.

AI Meeting Notes landed in May 2025 and takes transcription straight from system audio, launched on the Mac desktop app first, invoked from a page with /meet. It transcribes, summarizes, and extracts action items; you can take your own notes alongside while it records. Enterprise Search arrived in the same release: it queries Notion and connected tools — Slack, Google Drive, GitHub and Jira were live at launch, with Microsoft Teams, SharePoint and OneDrive added and Gmail, Linear, Zendesk, Salesforce and Box arriving in stages — and returns answers and summaries rather than a list of links. Notion's partner announcement with Cerebras claims streamed results under 300 milliseconds; I have no way to audit that number and pass it on as their claim.

Then there is the 2026 chapter: Custom Agents. These are autonomous workers you describe in a short job description; they run on schedules or triggers across Notion, Slack, Mail, Calendar, Figma, Linear and custom MCP servers. This is where the pricing story gets a second meter. Custom Agents consume Notion credits — $10 per 1,000 monthly credits at time of writing, free through 3 May 2026 and billed from 4 May 2026, on Business and Enterprise only. The core AI features (Agent chat, Meeting Notes, Enterprise Search) stay included in the plan; the agents that run unattended do not.

My testing notes

In my testing I ran the workspace Q&A against a test workspace seeded with conflicting information — two project pages with different launch dates for the same release, plus a database of around 120 content rows. On specific lookups ("what did we decide about the pricing page copy?") it found the right page and cited it every time. On questions spanning multiple sources it was right roughly four times out of five, and when it was wrong it was wrong quietly: one answer blended a superseded draft date with the approved one and presented it with full confidence and a real citation. That is the failure mode I warn readers about — the citation is genuine, the synthesis is not.

The second thing I measured was bulk database work, because that is where the Agent claims to outscale a human. I asked it to categorize and summarize all 120 rows of the content tracker in one run. It completed the pass in a single run of a few minutes, correctly updating existing records rather than duplicating them, and got the category wrong on a handful of rows where the title was ambiguous. Formula generation was the pleasant surprise: I described five property-logic requirements in plain English and four of the formulas were correct on the first attempt; the fifth, which depended on empty relation values, needed manual fixing.

Third, I used AI Meeting Notes on a 30-minute recorded call. The action items were genuinely usable — owner, task, deadline in most cases — but speaker attribution was weak and I had to set custom instructions before the output matched how my team writes follow-ups. The workflow is fast (type /meet, talk, get a structured page), and it is the feature I could hand to a non-technical colleague without a briefing.

What it does not do

The limits are where this review earns its keep. Start with the trial: Free and Plus workspaces get complimentary AI responses shared across the entire workspace, and the trial excludes AI connectors, bulk database autofill, and auto-updating autofilled properties. Once those shared responses run out, AI stops until somebody upgrades the workspace. Guests never get AI at all, on any plan.

Usage limits are the soft spot. Notion reserves the right to "temporarily reduce" your access to AI features depending on usage, with no published numeric ceiling I could find anywhere in the help center. I would like to tell you exactly how many Agent runs trigger that; the documentation does not say, and I will not invent a number. Third-party testing reports a context ceiling of around 50 pages per query scope, which matches what I saw when asking for summaries across large page sets — the answer came back noticeably thinner than the material I knew was in the workspace.

There is also no standalone API for the AI features. Notion has a public API for workspace content, and the Workers developer preview (beta) lets you build custom agent tools and external triggers, but you cannot call Notion AI programmatically the way you can call a model provider. If your automation lives outside Notion, the AI does not follow you there; the Custom Agents and connectors reach outward, but the center of gravity stays inside the workspace. And several headline features still carry beta labels — Enterprise Search, Research mode, Workers — which I weigh in the reliability column. Notion itself acknowledges prompt-injection risk for agents reading connected content and recommends restricting agent scope with permissions.

Finally, the data question: Notion states it does not train on your content, and Enterprise plans get zero data retention with model providers; everyone else keeps 30 days of retention. For most teams that is fine. For the compliance-driven buyer, it is a difference between Business and a custom-priced Enterprise contract.

Verdict

Rated 4.2 on the ToolSpotter rubric — depth of core functionality scores highest of the five axes at 40% weight, with value for money the drag: paying Business-tier seats for an entire workspace to serve two heavy AI users is the wrong shape for small teams, and the Custom Agents credit meter adds a bill that scales with ambition rather than headcount. What you get in return is unusually deep: an assistant with genuine access to the company's knowledge, meeting notes that produce usable follow-ups, an agent that edits databases at humanly awkward scale, and an ecosystem (connectors, MCP, a hundred-million-user install base) that will still be here next year. If your team already lives in Notion, this is the most coherent AI layer I have tested inside a workspace product, and the upgrade is easy to justify. If your knowledge lives in ten tools and your budget is one seat, buy a standalone assistant instead — Notion AI without Notion content is a very expensive cursor autocomplete.

Sources

• https://www.notion.com/pricing
• https://www.notion.com/pricing/ai
• https://www.notion.com/help/ai-pricing-and-usage
• https://www.notion.com/help/guides/using-notion-ai
• https://notion.com/releases/2025-09-18
• https://notion.com/releases/2025-05-13
• https://notion.com/blog/introducing-custom-agents
• https://techcrunch.com/2025/09/18/notion-launches-agents-for-data-analysis-and-task-automation
• https://techcrunch.com/2025/05/13/notion-takes-on-ai-notetakers-like-granola-with-its-own-transcription-feature/
• https://cerebras.ai/press-release/cerebras-enables-notion-to-deliver-real-time-enterprise-search-for-100-million-workspace-users
• https://en.wikipedia.org/wiki/Notion_(productivity_software)
• https://www.sentra.app/articles/notion-ai-pricing
• https://gotaskai.com/ai_tools/notion-ai/