Claude × Notion × GAS — Building an AI Operations Stack Without Writing Code

1 day ago
5 min read

For a non-programmer business professional looking to build an AI-powered operations stack from off-the-shelf cloud services, the best answer right now is Claude × Notion × GAS (Google Apps Script).

In the previous post,From Tacit Knowledge to Collective Intelligence — Using AI to Turn "What's in One Person's Head" into a Team's Greatest Asset, I wrote about why a mechanism for sharing tacit knowledge across a team matters. This post is about the how. We didn't arrive at this conclusion in a straight line — there was real trial and error along the way, and I want to lay out that path as well.

The first attempt: surely Notion alone would do it

Our first approach was to use Notion AI and keep everything inside Notion.

Notion was already functioning as our team's information backbone. Meeting notes, email records, tasks, client information — it was all there. The simplest answer seemed to be: just layer AI on top of what's already working.

The result fell well short of expectations.

Notion AI is fine for summarization and writing assistance, but it isn't built for compound tasks — reading context across multiple databases, making judgments, and updating other pages. The performance simply isn't there. Running the same workflows through Claude is, subjectively, several times faster and noticeably more accurate. At least at this point in time, Notion AI is a tool for assisting with individual tasks, not an engine that can autonomously run an entire operational workflow.

The breakthrough: Claude Cowork

In early 2026, Anthropic released Claude Cowork.

The moment I tried it, I thought, "This is it."

Claude Cowork is fundamentally different from the chat-based Claude. Triggered by schedules or events, it can execute multi-step tasks autonomously from start to finish. Access to local files, connections to external tools like Notion and Gmail, processing that involves conditional branching — all of this can run without a human in the loop.

If Claude Code is an autonomous agent for engineers, Cowork is an autonomous execution environment for the knowledge work of non-engineers. You don't need to write code to use it.

The core of the design: let each service do what it does best

The design philosophy we landed on is simple. Mind the cost-performance, and let each service do what it's best at. The division of roles looks like this:

Claude (judgment and generation). Handles compound reading, judgment, text generation, and page updates. That said, its ability to retain memory across conversations is limited, so we treat it as a volatile component.

Notion (memory and accumulation). Meeting notes, email records, tasks, workflow specifications — every piece of state and history is persisted here. It is the organization's source of truth and the core of our collective intelligence. We don't accumulate long-term memory inside Claude; we only give Claude pointers to where to look.

GAS (data ingestion). GAS is the input layer that feeds information from Gmail and Google Calendar into Notion. This means that by the time Claude starts processing, the information it needs is already organized in Notion. Claude only has to read Notion — it doesn't need to scan mailboxes or calendars directly. GAS absorbs the cost of data retrieval, freeing Claude's tokens for what AI is actually good at: judgment and generation.

You can have Claude generate the GAS scripts themselves. Describe what you want — "when this kind of email arrives, log it to this database in this format" — and the script comes out. Once it's running, it runs.

What's actually in production

Here are some of the workflows currently registered in Claude Cowork's scheduler and running on a daily cadence:

Hourly: Check the communications database in Notion and auto-generate reply drafts for unprocessed emails
Early every morning: Compile the previous day's meeting notes, today's meetings, and today's tasks into a morning summary page. By the time you wake up and open your laptop, it's already prepared
Every Saturday: Generate a weekly activity report to visualize progress
The 1st of each month: Generate a monthly review page

When I open my laptop in the morning, the email reply drafts, today's schedule, and the tasks I need to handle are all laid out in Notion. The "startup cost" of the day has dropped dramatically.

Why this works for a non-programmer

The essence of this stack is that the behavior of the AI is defined not in code, but in plain Notion text.

The instructions to Claude — which databases to read, which template to output, how to branch by condition — are all written in Japanese (or any other natural language) on Notion pages. What gets registered in Cowork is only the entry point: "read this Notion page and act on it."

If you want to change the workflow, you rewrite the Notion page. You don't touch code. You don't need to call an engineer.

The same goes for output quality. If Claude's output misses the mark, you write a comment on the Notion page pointing out what was wrong. On the next run, Claude reads that comment. The everyday act of "correcting" becomes, in itself, an improvement to the system.

The limits, and an honest assessment

Calling this combination "the best answer" right now comes with conditions:

It assumes you can operate without a programmer or internal IT department
It assumes the team is already using Google Workspace and Notion
It assumes a small team in scale

For environments with enterprise-grade information security requirements or complex integrations with internal systems, a different architecture is needed. This is strictly about what can be built within the range of off-the-shelf SaaS.

This system is also not a finished product. Improving the precision of individual skills, error handling, rolling it out across the team — these challenges remain. But on the direction, I have conviction.

In summary

Looking back at the order of our trial and error:

We tried to keep everything inside Notion with Notion AI — performance wasn't there. On compound tasks, both speed and accuracy are clearly inferior to a Claude-based setup
Claude Cowork launched — it can execute compound tasks autonomously, even when no human is present. What we had been looking for was here
We defined the role of GAS — by assigning the data ingestion layer to GAS, we reduced Claude's token load and improved both efficiency and accuracy of the overall flow

The Claude × Notion × GAS combination we ended up with is an AI-powered operations stack that a non-programmer business professional can build and operate within the range of commercially available cloud services. For now, this is our answer.

Envitalは、モビリティ・エネルギー分野を主なフィールドに、世界で生まれる先端技術や知見を最も必要とされる現場へつなぐことを生業としています。海外テクノロジー企業の日本市場参入支援から、日本企業の新規事業開発まで、国境や業界を越えた事業開発に取り組んでいます。