Blog Structure First, Then AI: Lessons From This Week's Notion AI Office Hour
Structure First, Then AI: Lessons From This Week's Notion AI Office Hour
A recap of this week's Notion AI Office Hour — the four levels of AI usage, why Notion is your memory layer, how MCP really works, and when a worker beats an agent.
Every week I run an open Notion AI Office Hour. No fixed agenda, livestreamed on YouTube (Building with Nasri), and steered entirely by whoever shows up. This week's was small but hands-on, and the questions did most of the driving. Here's what we covered, and the handful of ideas worth taking away even if you couldn't make it.
The four levels of AI usage
We opened with a quick poll asking people to place themselves on a simple AI maturity scale. It breaks down like this:
Level 1 — using AI just for chatting.
Level 2 — using AI for tasks and productivity.
Level 3 — running autonomous agents that work in the background.
Level 4 — a full business operating system powered by AI.
More than half of us sit at Levels 1–2, and only around 2% ever reach Level 4. The point isn't to rush to the top — it's to know where you are, and to take the next deliberate step.
Two types of agents, and no model lock-in
There's a useful line to draw between conversational agents (you chat, they act — the ChatGPT-style experience) and autonomous agents (they run quietly in the background while you sleep).
A detail people often miss: Notion agents are LLM-agnostic. There's no model lock-in, so you can reach for the heavy reasoning models when the task demands it and lighter ones when it doesn't.
Notion is your memory layer
Models like Claude don't hold persistent memory between conversations. So the framing I keep coming back to is this: treat Notion as the context and memory layer for your AI — structured databases it can read from and write back to, fast.
The practical lesson is counter-intuitive. Don't hand the AI your entire workspace. Deliberately limiting its context keeps responses faster and cheaper.
MCP is just a handshake
Whenever MCP comes up, I describe it as a bridge or a handshake — it gives an external AI the tools to read and update Notion. Connecting is simple (Settings → MCP → connect to Claude), and you don't need a premium plan for the connection itself.
But here's the catch we kept returning to: build your database structure in Notion first. Claude can't reliably design Notion structure on its own — it's far better at updating a container you've already shaped than inventing one from scratch. That was exactly the advice for one attendee's journaling workflow: build the journal database first, then let Claude fill it.
Workers vs agents — and keeping costs sane
We did a live build from plain natural language ("build me a task database"), then showed a favourite of mine: a movie watchlist where typing a title auto-fills the details. That's not AI burning tokens — it's a worker, a small script quietly pulling from a film API in the background.
That distinction is where the money lives:
Workers run as code and cost a fraction of an agent because they consume no AI tokens. Perfect for predictable, rules-based jobs.
Agents run on credits (300/month on the business plan, then about $10 per 1,000). A Canva image task, for reference, ran at 34 credits — roughly 34 cents.
The other lever is model selection: reserve the expensive models for creative and reasoning work, and use lighter ones for the boring steps like formatting or adding slugs.
Key takeaways
Structure first, then AI. Build the database in Notion before pointing Claude or any agent at it.
MCP is just a handshake. It grants permission and the means to read and write Notion — no premium plan needed for the connection.
Notion is your memory layer. Treat it as the AI's long-term context, and limit what you expose to stay fast and cheap.
Workers beat agents for repetitive jobs. If it's predictable and rules-based, a script costs a fraction of an agent.
Match the model to the job. Heavy models for reasoning; light models for the routine steps.
Give your agent skills. Reusable skills mean you stop repeating the same prompts and get consistent output every time.
Come to the next one
These sessions are free, live, and open — bring your questions or your half-built workspace and we'll work through it together. The full session is up on YouTube (Building with Nasri), and the open invite stands: connect with me on LinkedIn for a one-to-one.
How I turned a forgettable list of speaking gigs into a Netflix-style portfolio — with Notion as the single source of truth, AI doing the build, and a one-click path from page to live site.
Installing Notion Workers on Windows isn't plug-and-play yet. Here's the three-hour story of WSL2 setup, cross-OS gotchas, and the workaround that finally got my first Worker shipping live data into Notion.
A behind-the-scenes look at hosting the OpenClaw SG x Notion event — demoing Beaver, introducing Notion Custom Agents, and the unexpected conversations that made the night special.
Made with Bullet