Notion

Notion builds a collaborative workspace for documents, knowledge, and project management, with AI that can edit content, retrieve information, and take action across connected tools. Its AI products evolved from a writing assistant released broadly in February 2023 to a personal Agent in September 2025 and Custom Agents for recurring background work in February 2026. Pages and databases serve both as the workspace people use and as context, memory, and coordination tools for agents.[3][6][7][2]

According to co-founder and CEO Ivan Zhao, Notion received early access to GPT-4 in fall 2022. At a company offsite, Zhao, co-founder Simon Last, and members of the product and engineering teams developed an AI prototype in a hotel room. Zhao and Last extended their stay by three days to finish it. Notion launched a private alpha in November 2022, less than a month after the offsite.[4]

The team began with writing assistance and gave preset prompts and custom text input equal weight: presets supplied an accessible starting point, while custom prompts let users adapt AI to their needs. It chose the spacebar as a new entry point despite early feedback that it was unintuitive. Zhao also recalled deciding to enable AI by default to avoid developing separate versions of Notion, with an Enterprise option to turn it off. Availability did not mean automatic activation: the February 2023 launch required a user action before generating content or sending requests to AI providers.[4][3]

The February 22, 2023 general release followed ten weeks of invite-only testing and nearly two million waitlist signups. Zhao reported that highlighting existing text and asking for help was the leading interaction, and that users returned most consistently to “Improve writing.” Notion redesigned the interface around editing existing work, iterative follow-up prompts, and suggested actions suited to the user's context. At launch, AI was a $10-per-member monthly add-on; the Writing Suite's stated policy prohibited using customer data to train Notion's or its partners' models.[3]

The underlying workspace uses blocks as a shared abstraction. Paragraphs, bullets, and pages are blocks, and databases collect pages. A paragraph dragged into a database can become a page; a page pulled out can become a link. In his June 2023 explanation, Notion AI engineer Linus Lee described this fluidity as the basis for combining documentation and workflows in one connected workspace.[1]

By June 2023, Notion AI had three principal surfaces: Writer, AI Blocks, and Autofill. Writer drafted, summarized, and edited text. AI Blocks confined output to a recognizable area of a page; Lee's team used them in meeting templates to regenerate summaries and action items at the press of a button. Autofill extracted information into database columns—for example, pain points or competing products mentioned in user interviews—and could update a property when its underlying page or row changed. Unlike Writer and AI Blocks, it could operate automatically in the background.[1]

Lee described multilingual few-shot examples as a way to keep English-language instructions from producing English answers to documents in other languages. The team checked prompt changes against handwritten regression cases and voluntarily shared user feedback, using language models to judge properties such as summary quality and response language. In one Autofill prototype test, Lee recalled summarizing approximately 6,000 pages and consuming roughly two or three million tokens in about 20 minutes. In November 2023, Notion added Q&A to answer questions using information across a workspace.[1][4]

Notion 3.0, launched September 18, 2025, introduced a personal Agent that could create pages and databases, search connected tools, and execute multi-step workflows. Notion advertised up to 20 minutes of autonomous work across hundreds of pages at once. Users could personalize behavior through an editable instructions page. Custom Agents were still forthcoming at this launch.[6]

In an April 2026 interview, Last recalled rebuilding the agent effort four or five times after experiments beginning in late 2022. Early approaches included a JavaScript coding agent and a custom XML representation that mapped losslessly to Notion blocks. The XML suited Notion's internal structure but was unfamiliar to the models, so the team developed Notion-flavored Markdown and adopted SQL-like database querying. Head of AI Sarah Sachs described a later move from shared few-shot prompts toward tool definitions, allowing product teams to own individual tools with evaluation support.[2]

Last and Sachs attributed the long development period to limited model reliability, short context windows, and the demands of background operation. Permission design also required repeated attempts: a shared agent might appear in a Slack channel whose audience differed from the audience permitted to read its documents. Last said background agents received explicitly granted permissions and could not edit those permissions themselves. Repairing an agent used a synchronous administrator interaction with confirmation. The team delayed launch by roughly a month to make chat the main interface for both configuring and using an agent, with settings available alongside it.[2]

Custom Agents entered public beta on Business and Enterprise plans on February 24, 2026. They run on schedules or event triggers, handling work such as Slack Q&A, task routing, and recurring reports across Notion and connected tools. The launch included run logs, reversible changes, detailed permissions, usage dashboards, and automatic pausing at credit limits. Notion reported that early testers had built more than 21,000 agents. Its customer examples included Ramp's Product Oracle for roadmap questions, Braintrust's Deal Spotter for upgrade opportunities, and Clay's Incident Reporter for post-mortems.[7]

On May 4, 2026, Notion announced that Custom Agents were out of beta and that users had created more than one million during the beta. Usage-based pricing began that day, with enforcement at each workspace's next monthly service date on or after May 4. The announced recurring price was $10 per 1,000 Notion Credits. Sachs explained that credits abstracted over model choice, token usage, serving tiers, and differently priced services such as web search. Creation counts describe agents built, rather than active workflows or unique customers.[8][2]

By April 2026, Sachs described three distinct evaluation purposes: regression tests in continuous integration, launch-quality report cards for user journeys, and frontier or headroom evaluations deliberately targeting difficult tasks at roughly a 30% pass rate. Teams owned their evaluations while a platform team maintained the framework. Model Behavior Engineers formed a separate career path combining evaluation design, failure analysis, and qualitative understanding of model behavior; Sachs emphasized that an engineering background was not required.[2]

Sachs said foundation-model training need not be Notion's core competency, while describing targeted fine-tuning and investment in retrieval, ranking, and query generation. She reported that agents generated the majority of search traffic on AI-enabled Business and Enterprise plans. This changed retrieval priorities: inclusion in a useful set of results mattered more than the precise ordering optimized for human clicks, and snippets needed to suit agent consumption. She also described Meeting Notes as data capture that supplies collaboration context to search and agents.[2]

Sachs reported about 50 people in core AI capabilities and infrastructure, plus another 30–40 in teams packaging those capabilities into chat, Custom Agents, and Meeting Notes. Product teams also owned the corresponding agent tools. She credited designer Brian Levin with promoting “demos over memos” and Jimmy on the database collections team with initiating image generation, supported by resources for working directly with Gemini and evaluating the feature.[2]

The Developer Platform announcement on May 13, 2026 introduced Workers, Notion's hosted sandboxed runtime for custom code. Workers could synchronize external data, provide custom agent tools, and handle webhook triggers, giving workflows explicit code-based logic. Workers launched in public beta on Business and Enterprise plans; the ntn CLI became available on all plans for reading and writing, while deploying and managing Workers required Business or Enterprise. External Agents and the External Agent API were offered through a private-beta waitlist, with named partner agents including Claude Code, Cursor, Codex, and Decagon.[9]

The September 15, 2026 Notion 3.7 release added reusable team skills and shared skill libraries. Instructions maintained in Notion could be exported as a SKILL.md file with approved supporting files for use in other agents. Custom Agents could call other Custom Agents as sub-agents with their own instructions, context, access, and models. They could also receive meeting transcripts and summaries to perform automatic follow-up work. The release announced an Agent SDK public beta on Business and Enterprise and the Notion Knowledge Board for evaluating models on everyday work.[13]

The pages and databases used by these agents preceded Notion AI. Notion's earlier product history supplies the context for the workspace those agents now operate in.[6]

Zhao's September 2024 retrospective dates Notion's start to 2013 and its first signs of traction to the 2018 Notion 2.0 Product Hunt launch. He recalled four early product rebuilds, including a rewrite in Japan with Last after an unsuitable technology stack. His 2023 account places the company's near-failure and move to Kyoto in 2015. These early product rebuilds preceded the later agent rebuilds. Notion describes 1.0 as a collaborative canvas for documents and knowledge, and 2.0 as adding databases and integrations.[5][4][6]

In October 2021, Notion announced a $275 million fundraise from Sequoia Capital and Coatue Management at a $10 billion valuation. The same announcement introduced a Japanese-language product beta and Japanese customer support, alongside a Tokyo office expansion led by Regional General Manager Katsukiyo Nishi. Notion had opened an EMEA headquarters in Dublin earlier that year. The company reported that 80% of its users were international and that its Japanese daily active user base had grown 500% over the preceding year.[10]

On September 3, 2024, Zhao reported that Notion had passed 100 million users the previous month, recalling the one-million-user milestone in 2020. The announcement did not define the 100 million as active users or paying customers.[5]

Notion's acquisition of Skiff had direct consequences for that service's users. Skiff's migration notice states that Mail, Pages, Calendar, and Drive were discontinued on August 9, 2024. Accounts and data were not converted into Notion accounts; previously enabled automatic mail forwarding continued only until February 9, 2025.[12]

Notion launched Calendar in 2024 and introduced Notion Mail on April 15, 2025. Mail combined AI auto-labeling with configurable inbox views, reusable snippets, and Calendar scheduling through /schedule. Its writing assistant could draft follow-up emails using Notion documents as context, extending AI assistance into another daily work surface.[11]