Notion

AI inside the workspace

Notion makes workspace software built around documents, databases and collaboration. Its AI features use those existing structures as places to read information and do work. In the September 2025 agent release, users could choose underlying models including Claude Sonnet 4 and GPT-5. Notion's contribution was the application around them: workspace context, permissions and actions on the pages people already used. [1]

Starting with the words already on the page

Notion AI became publicly available on February 22, 2023, following an invite-only alpha introduced the previous November. It could summarize notes, improve writing and generate new text inside the editor. The launch removed the waitlist and made AI available as an add-on across Notion plans; public availability did not mean unlimited free use. [2]

The company said its alpha feedback changed the product. Users more often selected their own text for help than asked the assistant to create whole articles from a blank page, and improving existing writing was a recurring use. This is a useful early example of generative AI entering an established workflow through editing and summarization, with that usage pattern reported by the vendor itself. [2]

At Claude's introduction on March 14, 2023, Anthropic named Notion as a partner from its closed alpha. Notion's co-founder Akshay Kothari described Claude's writing and summarization capabilities contributing to Notion AI. The announcement establishes an actual product integration and launch partnership; it does not establish that Claude was the only model behind Notion AI. [3]

From editing text to taking actions

Notion 3.0, released on September 18, 2025, extended the product to personal agents that could create documents, build databases and carry out work across multiple pages. Instructions could live in a Notion page, and connected tools supplied additional context within the user's permissions. The release thus expanded the scope from helping with an answer or paragraph to changing workspace objects over a sequence of steps. [1]

The same announcement previewed shared Custom Agents running on schedules or triggers, but labeled them as coming soon. Its broad claims about what agents could accomplish and the time they could work were product claims, not an independent reliability study. Keeping those boundaries visible distinguishes a shipped personal agent from a promised autonomous workflow. [1]