Notion AI

Notion AI is Notion’s suite of AI tools for writing, finding information and automating work inside its connected workspace. Introduced as a writing assistant in November 2022, it expanded into workspace question answering, meeting transcription and research before Notion rebuilt it around Agents in September 2025. Custom Agents followed in February 2026, adding background workflows driven by schedules and triggers; shared agent skills arrived in September 2026.[4][6][7][8][9][12]

The AI effort grew inside an existing workspace built around composable pages, blocks and databases. Its earlier history explains the software structures that later became places for AI to write, store memory and coordinate work.[1][2]

Notion started in 2013. In his 2024 retrospective, co-founder Ivan Zhao recalled four rebuilds during its early “lost years,” including a rewrite in Japan with Simon Last after they had reached the right product on the wrong technology stack. Zhao dated the first signs of traction to the 2018 Notion 2.0 Product Hunt launch. Latent Space’s June 2023 account places the Japan rewrite in Kyoto in 2015, when the founders laid off the rest of the team and discarded their previous app.[10][1]

Notion’s core abstraction is the block: a paragraph, bullet or page can all be blocks, and a database is a collection of pages. In June 2023, Notion AI engineer Linus Lee explained that a paragraph could be dragged into a database to become a page, while a database page could be pulled out as a link. This shared structure lets documents and workflows coexist in one workspace. For Lee, a central question for Notion AI was what an equally useful building block for AI would look like.[1]

Latent Space’s June 2023 introduction credits Zhao and Last with building the early Notion AI prototype in a hotel room during Notion’s 2022 company retreat. On November 16, 2022, Zhao announced a waitlist-only private alpha offering drafting, brainstorming, editing, translation and summarization inside Notion. He connected the product to Notion’s mission to make software toolmaking ubiquitous and cited Doug Engelbart’s vision of computers augmenting human intellect as an inspiration for building Notion.[1][4]

The alpha also attracted external experimentation. In a December 28, 2022 account, swyx described using prompt injection to elicit feature prompts within two hours of gaining access. The apparent recovery was incomplete: the article records Jake from Notion’s objection that some text was verbatim, some reordered or invented, and much missing. The author ultimately characterized the recovered prompts as educated guesses.[3]

Notion AI became available to everyone on February 22, 2023. Notion reported nearly two million waitlist signups and an important difference between its expectations and alpha usage: people most often highlighted existing text for help, and “Improve writing” was the feature they returned to most consistently. Standalone drafting prompts were infrequently used when alpha users created new pages. Notion redesigned the interface around working with existing words, follow-up prompts and suggested actions appropriate to the user’s location in a document. The launch offered AI as a $10-per-member monthly add-on, with a 20% discount for annual billing.[5]

Lee joined the AI team in January 2023 after taking 2022 off to explore how text interfaces could evolve. He said Notion offered both room to prototype quickly and an established user base from which to learn. His design framework extended beyond generating prose: knowledge work also involves understanding, synthesis and proposing actions. As he put it in the June 2023 interview, “So much of the value of what humans sit down and do at work isn't actually in the words that they write.”[1]

By June 2023, Lee described three product cornerstones: Writer, AI Blocks and Autofill. Writer helped draft, summarize and edit documents. AI Blocks provided a defined area that could be refreshed without prompting from scratch—for example, the AI team’s meeting-note templates had a block that regenerated a summary and action items when someone pressed a button. Autofill brought AI to database columns, extracting information such as pain points or competing products from user-interview pages. Unlike Writer and Blocks, configured Autofill properties could update in the background when a page or row changed.[1]

Lee argued that suggested actions and constrained interfaces could steer users toward model strengths and make expectations clearer. The team nevertheless retained custom prompts for power users and for discovering new use cases. Blocks and autofilled cells also made the AI’s writing area visible within a collaborative document. Automation increased the potential workload: Lee recalled a prototype test summarizing roughly 6,000 pages and consuming approximately two to three million tokens in about 20 minutes.[1]

On November 14, 2023, Notion released Q&A in beta, extending AI from individual documents to questions answered using information across workspace pages and databases. Zhao credited close partnerships with Anthropic and OpenAI. The announcement said Q&A referenced only pages the user had permission to view and that customer data was not used to train the AI models.[6]

Notion AI for Work, announced May 13, 2025, combined AI Meeting Notes, Enterprise Search and Research Mode. Meeting Notes transcribed conversations, produced summaries and action items, and stored searchable records connected to projects. Enterprise Search extended question answering to connected apps, while Research Mode used workspace sources and the web to draft documents such as project updates and research reports. Meeting Notes and Research Mode were already available at the announcement; Enterprise Search and connectors were rolling out. Notion introduced all-in-one packaging starting at $20 per user per month.[7]

Notion 3.0, released September 18, 2025, rebuilt Notion AI as Agents. The personal Agent could build a launch plan, break it into assigned tasks and draft documents across multiple pages, drawing context from the workspace, connected tools and the web within the user’s permissions. Instructions and memory used Notion pages and databases. Notion advertised more than 20 minutes of multi-step actions. Custom Agents were still forthcoming at this release.[8]

Custom Agents launched on February 24, 2026, under Notion 3.3. Teams could create and share agents that ran from schedules or triggers to answer recurring questions, route requests and generate reports. The release described controlled access, run logs and visible, reversible changes. Notion reported that early testers had created over 21,000 Custom Agents and that 2,800 agents were working internally at Notion. It announced a free trial through May 3 and a planned transition to Notion credits on May 4, as an add-on for Business and Enterprise plans.[9]

In their April 2026 interview, Last and Sarah Sachs described a consequential setup decision: they changed Custom Agents from a settings-first interface to a chat-first interface in which the agent could help configure and debug itself. Last recalled delaying launch by roughly a month to make the change; Sachs credited three engineers from three teams with completing it. Last distinguished that supervised setup from background execution: agents had no permissions by default, required explicit grants and could not edit their own permissions. Fixes could instead enter a synchronous administrative interaction where the user could see the proposed changes.[2]

Last recalled rebuilding the agent effort four or five times. The early version used JavaScript APIs and generated code, but the models were not good enough at coding. A subsequent custom XML format mapped cleanly to Notion blocks but was unfamiliar to the models. The team moved to Notion-flavored Markdown and a SQL-like database-query interface, favoring representations models already understood over exposing Notion’s internal complexity. Sachs also credited frontier labs and Fireworks with early attempts to fine-tune function calling before standard tool calling was available.[2]

Sachs described a later shift from centralized few-shot prompt strings to tool definitions owned by product teams. Previously, changing examples meant editing one shared string, with interactions between example selection and ordering; only five or six people could approve changes. Tool definitions and evaluations allowed ownership to spread across the company. In April 2026, Last described progressive tool disclosure as another update nearing shipment: showing every tool at once added token cost and could cause a niche tool to interfere with the agent’s wider behavior.[2]

Evaluation evolved alongside the products. In June 2023, Lee described multilingual few-shot examples to keep outputs in the appropriate language, handwritten regression cases, voluntarily shared user feedback and language-model judges for summaries and language adherence. By April 2026, Sachs distinguished regression tests, launch-quality evaluations of user journeys and frontier or headroom evaluations deliberately targeting about a 30% pass rate to expose capabilities still beyond the models. Product teams owned their evaluations while the central team maintained the framework. Braintrust’s customer account describes a complementary practice: searching traces for language-adherence failures and turning them into targeted evaluation datasets.[1][2][11]

Rapidly changing tools shaped the team’s training choices. Sachs and Last said tool-specific fine-tuning could slow development because each tool change might require retraining and shipping a new model; they emphasized improving tools, the harness and verification instead. This did not exclude training: Sachs described selected capability fine-tuning and investment in retrieval and ranking. She reported that most search traffic on Business and Enterprise AI-enabled plans came from agents, prompting changes to query generation, snippets and retrieval objectives. Agent retrieval placed different demands on the returned set of results than human click-through ordering did.[2]

Pages and databases also supplied coordination and memory. Last described agents writing to shared databases that other agents watched, and giving an agent an editable page as memory. Sachs’s internal meeting workflow illustrated how the pieces connected: an agent prepared a standup pre-read from Slack and GitHub, Meeting Notes captured the discussion, and a triggered agent filed tasks and sent agreed Slack follow-ups. For pricing, Sachs explained credits as an abstraction over model-dependent token costs, GPU serving and web search, allowing expensive background workflows to be offered through usage-based charging.[2]

Sachs described approximately 50 people working on core AI capabilities and infrastructure in April 2026, with another 30–40 in partner teams packaging capabilities into chat, Custom Agents and Meeting Notes. Product teams also owned the tools through which agents interacted with their surfaces. The parent workspace had a much larger reported user base: Zhao announced in September 2024 that Notion had passed 100 million users in August 2024, a Notion-wide milestone rather than an AI-user count.[2][10]

Notion 3.7, released September 15, 2026, added reusable team instructions called skills, organized and shared through Notion databases. Users could invoke a skill in Agent chat or export it as a SKILL.md file with approved supporting files for Claude Code, Codex, Cursor, Gemini or Grok. The release also enabled Custom Agents to call other Custom Agents as sub-agents with their own instructions, context, access and models, and announced an Agent SDK in public beta for Business and Enterprise plans.[12]