Notion: Claude-powered agents inside the workspace cut customers' information search time by 35%

• Notion builds agents on Claude (customer story) — Anthropic / Claude (кейс-стади)
• Notion Q&A | Claude Managed Agents (customer story) — Anthropic / Claude (кейс-стади)
• Notion 3.0: Agents (release notes) — Notion (официальные release notes), 2025-09-18

Background

Notion is an all-in-one workspace platform: documents, knowledge bases, tasks, and wikis for teams. The company was among the first on the market to embed Claude into its product and has kept scaling the AI layer since: from text autocomplete and knowledge-base Q&A to full agents that execute multi-step tasks right inside the workspace.

The turning point was the Notion 3.0 release on September 18, 2025: Notion AI was rebuilt from the ground up around agents. Per the release notes, the agent executes any action available to a user in Notion, works across multiple pages at once, sustains over 20 minutes of multi-step actions, updates or creates hundreds of pages at a time, and pulls context from connected tools — Slack, Google Drive, GitHub — within the user's permissions. The agent has memory and custom instructions: over time it remembers writing style, preferred formats, and filing rules — "like a coworker that's been around and has genuine context."

Anthropic's case study describes the current generation of this integration: Notion agents on Claude Opus 4.6 — the first time an Opus-tier model became available inside Notion Agent — and the Claude Managed Agents infrastructure, cloud-hosted composable APIs for building and deploying agents at scale. Notion product manager Eric Liu explains the choice in words rarely seen in enterprise case studies: "Claude powers many use cases where we really care about the response's feel and vibe." Beyond tone, the team cites reasoning quality and instruction-following as selection reasons.

For Anthropic this is a showcase on two lines at once: as an example of a product built on Claude, and as a reference for the Managed Agents infrastructure it sells to other agent builders.

Problem

Knowledge work drowns in busywork: employees spend time searching knowledge bases, reassembling the same content, and doing routine document operations. Notion co-founder Simon Last, who leads AI, states the project's mission in one line: "We want to automate busy work out of knowledge work and free up humans so they can do more high-value activities."

The product signal was unambiguous: Notion saw customers willing to jump through hoops to get a native agent experience inside Notion — and Claude was the model they most often wanted. An external chatbot that requires copying context loses to an embedded agent that already sees the team's documents, decisions, and tasks.

It is worth noting why busywork is the right first target for agents (our editorial observation, not the case study's claim): knowledge-base search, content reassembly, and routine document operations are tasks with verifiable outcomes and a low cost of error. A wrongly retrieved document annoys but doesn't wreck a process — unlike, say, autonomous actions in financial systems. That makes knowledge routine the ideal proving ground for getting teams accustomed to agents.

The technical challenge is non-trivial. An agent is not a one-shot model call: it needs long sessions (a task can run from 20 minutes to hours), memory management between steps, and parallelism — users want to launch dozens of tasks simultaneously, not wait for one. All of it collides with economics: long agent sessions with large context cost orders of magnitude more than short queries, and without inference optimization such a feature wrecks the product's unit economics. An extra requirement is first-try quality: if an agent's output can't be shown to a colleague or client immediately, the time saved turns into a rework loop.

Solution

Notion built its agents on Claude Opus 4.6 on top of Claude Managed Agents — Anthropic's cloud infrastructure providing long-running sessions, memory management, and prompt caching out of the box. "We integrated Claude Managed Agents... to make that possible," says Eric Liu. AI lead engineer Sarah Sachs adds about the model: "Opus 4.6 excels at interpreting what users actually want, producing shareable content on the first try."

The central interface is a task board. A user assigns work to an agent the way they would to a colleague: from a single board, agents run 30+ concurrent tasks. "You're not limited to one task. For our customers, what that means is that they can kick off a ton of jobs in Notion, 30 or 40 at the same time," Liu explains. The agent picks up the task's full context — connected pages, design systems, API docs, product requirements — and reports progress as it works. The interaction follows human conventions: version history, suggested edits. "You're working with Claude just like a colleague," Liu puts it.

The task range spans code generation and prototyping to artifacts: presentations, PDFs, websites, spreadsheets, and client deliverables. A distinctive mechanism is the self-improving skills database: after every completed task or merged pull request, the agent updates its own knowledge base of how things get done on this team — the system gets more useful with use, without manual training.

The economics were solved with prompt caching: it cut inference costs by 90% and latency by up to 85%. That is not cosmetics but the feature's existence condition: without caching, long agent sessions with repeated workspace context would cost ten times more, and agents would have to be strictly rationed.

The platform frame matters too: per the Notion 3.0 release, agents operate within the user's existing permissions, connect to external sources via MCP connectors and AI connectors (Notion Mail first, Box and Outlook announced), help generate database formulas, and custom agents can run autonomously on scheduled triggers for specific jobs — customer feedback processing, project updates, IT requests. Database-scale operation is a capability of its own: the agent updates or creates hundreds of pages in one operation, turning it from a "text writer" into a tool for bulk changes across the whole workspace.

Notion remains a multi-model platform: the 3.0 release announces access to several models, including OpenAI's, with no per-model fees. But the agent core described in Anthropic's case study runs on Claude — and the very fact that a platform with a choice put its flagship scenario on Opus 4.6 is, predictably, used by Anthropic as an argument.

Result

The impact is measured primarily on the side of Notion's customers. Osaka Gas cut information search time by 35%. Remote saves 10 minutes per search across 300 daily queries. dbt Labs saved over $35k a year by dropping separate AI tools — Notion's agents covered those scenarios. During onboarding, new employees query the AI assistant 10–20 times a day in their first weeks. Inside Notion itself, one recorded example saw 12 hours of prototyping work collapse into about 20 minutes: "Then your whole team can jump in and refine it together," says Liu.

For Notion itself, the key result is economics: prompt caching cut the cost of running agents by 90% and latency by up to 85%. That is what makes agents sellable as a mass feature rather than a premium option for the few.

Framing. The case is published by Anthropic — the vendor selling both the model and the Managed Agents infrastructure; customer metrics (35% at Osaka Gas, $35k at dbt Labs) come without measurement methodology; 12 hours → 20 minutes is a single example, not an average; and the before/after comparison does not isolate Opus 4.6's contribution from the general evolution of Notion AI.

Editorial analysis. First: this is one of the few cases where a vendor story contains third-party metrics — the customer's customers (Osaka Gas, Remote, dbt Labs), and such numbers are more convincing than internal benchmarks, though they too passed through Anthropic's marketing filter. Second: the pairing of "90% savings from caching" with "30+ concurrent tasks" reveals the real cost structure of agent products — the winner is not whoever has the smarter model but whoever learned to reuse context; for anyone building agents into their SaaS, that is the case's most transferable takeaway. Third: the dbt Labs story ($35k saved by dropping third-party AI tools) sketches the market's consolidation dynamic — platforms' built-in agents are eating the budgets of specialized AI add-ons, worth remembering when deciding what to build yourself versus what comes bundled with the platform. A fourth observation concerns the pace of autonomy growth: the Notion 3.0 release in September 2025 claimed "over 20 minutes of multi-step actions," while Anthropic's case study already describes tasks running from 20 minutes to hours. In a matter of months, the agent's autonomous horizon grew severalfold — and it is this curve, not any single metric, that determines which categories of busywork agents will take over next.

Lessons learned

1. Measure an AI feature by your customers' metrics, not just your own: 35% time savings at Osaka Gas and $35k/year at dbt Labs beat any internal benchmark.
2. Prompt caching is not a 'later' optimization but a viability condition for in-product agents: 90% inference savings change the feature's unit economics.
3. Native wins: users tolerate friction to get agents inside their everyday tool — an embedded agent beats an external chatbot.
4. Agents need infrastructure, not just a model: long sessions, memory, and parallelism (30+ tasks from one board) are a platform layer; Notion bought it ready-made (Managed Agents) instead of a year of in-house work.
5. Design agent interaction around human conventions — a task board, version history, suggested edits: it lowers the trust barrier ('working with Claude just like a colleague').
6. A self-improving skills database is a compounding mechanic: an agent that records 'how things are done here' after every task grows more valuable with each day of use.
7. Onboarding is an underrated ROI scenario: 10–20 daily queries from a new hire mean the agent effectively replaces part of mentoring.

Frequently asked questions