Ivan Zhao’s most consistent product thesis is that software should give non-programmers composable building blocks rather than force every team into a fixed workflow. Notion’s move into AI extends that thesis: the workspace is no longer only where people write and organize information, but a system that can retrieve, summarize, generate, and increasingly act over that information. The opportunity is substantial; so are the permission, provenance, and reliability requirements.

As of September 13, 2026, Notion identifies Zhao as co-founder and CEO. In his own account of the product’s development, he says the company moved to Kyoto after it almost failed in 2015 and rebuilt the application. That claim appears in Zhao’s Notion AI retrospective, so it can be attributed to him. It does not justify a dramatized scene about a particular house, conversation, or emotional state that the source does not document.

The short answer: Notion is trying to make knowledge executable

Notion began with flexible pages, databases, and blocks. AI changes the product from a toolkit users explicitly assemble to one that can infer what information matters and generate an answer or action. This creates three overlapping products:

  1. A collaborative source of record for documents, projects, and structured data.
  2. A retrieval layer across workspace content and connected applications.
  3. An AI work surface for writing, analysis, meeting notes, databases, and delegated tasks.

The strategy is coherent because the workspace already contains context. It is also risky because an error can look more authoritative when it is written in the same interface as the underlying company knowledge.

Zhao’s product philosophy is visible in attributable material

Notion’s First Block conversation records Zhao and co-founder Simon Last discussing the founding story, early product-market fit, a small team, horizontal product scope, and the future of AI. The source is produced by Notion, so it is useful for understanding the founders’ stated intent rather than independent proof of business performance.

That intent centers on giving users primitives—text, databases, relationships, views, and permissions—that can be combined into their own systems. The advantage is adaptability. The cost is that blank-canvas products can require setup, governance, and user education. Templates and AI can reduce that friction, but they can also hide assumptions about how a team should work.

The right product question is therefore not whether Notion can replace every specialized application. It is where a shared, configurable layer reduces handoffs without weakening the domain controls that specialized systems provide.

Notion AI moved from writing toward retrieval

Zhao says Notion launched an AI writing alpha in 2022 and later added Q&A over workspace information. Notion’s Q&A announcement describes answers grounded in pages, databases, projects, and meeting notes. These are company descriptions of intended behavior. They are not guarantees that every answer is complete or correctly cited.

Retrieval over an enterprise workspace has a different failure profile from generic chat. A response can omit the latest policy, blend two projects, expose a page through a misconfigured permission, or present a draft as approved guidance. Evaluation should use versioned company questions with known answers and permission boundaries.

For each answer, a useful system should show the source page, relevant passage, page owner, and last update. If evidence conflicts, it should expose the conflict rather than silently choose the most fluent text.

The privacy claims need plan-specific reading

Notion’s current AI security and privacy documentation says AI features honor existing workspace permissions and that Notion and its AI subprocessors do not use customer data to train models by default. It also describes different provider-retention defaults: zero retention for Enterprise workspaces and up to 30 days for non-Enterprise workspaces, subject to feature-specific settings and exceptions.

These are detailed vendor commitments, but buyers should verify the contract, plan, enabled features, subprocessor list, and region that apply to their deployment. “Not used for training” is not the same as “not transmitted,” “not retained,” or “end-to-end encrypted.” Embeddings, connectors, backups, logs, and support access have separate lifecycles.

An enterprise review should map:

  • which content is sent to which model or vector provider;
  • how connector permissions are synchronized and revoked;
  • whether deleted pages remain in indexes, backups, or generated artifacts;
  • which administrators can enable web search or external agents;
  • how prompt, source, and tool logs are retained;
  • whether users can export evidence for an important answer.

The NIST Generative AI Profile is a useful external checklist for confabulation, privacy, information security, and human overreliance. It does not certify Notion.

Business scale should not substitute for product evidence

A 2024 Forbes profile reported company and valuation estimates alongside figures attributed to Zhao. Those data points describe a moment and mix company statements with Forbes estimates. They do not establish current revenue, ownership, profitability, or Zhao’s net worth, and this article does not repeat those personal estimates.

For buyers, Notion’s adoption is relevant only insofar as it supports service continuity and ecosystem depth. It cannot answer whether the product fits a regulated workflow, whether retrieval quality is sufficient, or whether consolidation will lower total cost.

A practical acceptance test

Test Notion AI on a bounded workspace before connecting the whole company. Include:

  1. Questions with one current authoritative source.
  2. Questions where an old page conflicts with a new decision.
  3. Users who lack access to one relevant page.
  4. Ambiguous names shared by projects or customers.
  5. Requests that would require an external action or confirmation.
  6. Deleted, archived, and recently moved content.

Measure source precision, answer completeness, unauthorized disclosure, abstention, review time, and correction latency. For agent-like actions, enforce approval and least privilege outside the model. A conversational request is not an access control.

What remains unknown

Public sources do not reveal every Notion model provider, internal evaluation result, customer retention figure, private board discussion, or Zhao ownership percentage. They also cannot support claims about his personal motives or recreate undocumented moments from the company’s early years.

The defensible conclusion is that Zhao’s building-block philosophy gave Notion a strong substrate for contextual AI. The same concentration of company knowledge that makes the product useful raises the cost of a permission or provenance failure. Notion’s next advantage will be earned through trustworthy retrieval and controlled action, not by claiming to be an everything app.

Source and correction note

This revision removes an invented Kyoto opening scene, personal-wealth estimates, unsupported internal metrics, and speculation about Zhao’s private views. Founding and product claims are attributed to Zhao or Notion; independent reporting is identified separately. Documentation was checked through September 13, 2026.