AI work should not hide in chat. Put it on a Notion board

AI work disappears too easily inside private chats. One person asks Claude to summarize a document. Someone else lets Cursor take a first pass at a text or a script. A third person pastes meeting notes into a chatbot and gets a task list back. Useful, yes. Easy to follow up? Not really.

Notion's 3.6 release points in a more practical direction: AI agents can work from the same surface as the rest of the team. The release introduces External Agents starting with Claude and Cursor, speaker labels in AI Meeting Notes, more Microsoft file support, Outlook connections, HTML blocks, and more MCP connections for Custom Agents. For a smaller team, the interesting part is not another AI button. It is that AI work can become visible work cards.

Source: Notion 3.6: External Agents, HTML blocks, and more

An AI agent is an AI workflow that can take a task, use sources or tools, and return an output. That output might be a chat answer, but it might also be a document draft, a calendar action, a customer note, or a file that needs review. When that work lives in a private chat, the rest of the team only sees the final result, if they see anything at all.

That is where an agent board helps.

What Notion 3.6 changes for everyday work

Notion says External Agents can be assigned tasks from a shared board, @-mentioned like teammates, and watched while they run. The same release mentions speaker labels in AI Meeting Notes, support for file types like PPTX, XLSX and DOCX, Outlook Mail and Calendar, interactive HTML blocks, and more MCP connections for Custom Agents.

For a Nordic small business, school office, course provider or solo operator, this makes AI feel less like a separate tab and more like part of the work itself. A quote request, meeting, course evaluation or customer question can become a card where everyone can see:

• what AI should do
• which sources it may use
• which tool or agent is running
• who reviews the output
• what must not be sent, booked, changed or published without approval

That sounds less dramatic than "autonomous agents". Good. It is also much easier to use.

Build the agent board before connecting more tools

Do not start with "which agent should we buy?" Start with "where will the AI work be visible?"

A simple Notion board is enough. Call it AI work and use five statuses:

• Inbox: tasks someone wants to try with AI.
• Ready for AI: the task has a source, outcome, owner, and stop rule.
• Running: the agent or person is working on it.
• Needs review: AI has produced a draft, but nothing is approved yet.
• Done: a human reviewed it, changed it and decided the next step.

Each card needs a few fixed fields. Not twenty. Just enough to avoid confusion:

• Desired output: for example "customer reply draft", "meeting summary" or "course email draft".
• Sources: links, files, meeting notes or database pages the AI may use.
• Agent or tool: Notion Agent, Claude, Cursor, or a person running the prompt manually.
• Allowed actions: read, summarize, draft, compare, suggest.
• Requires approval: send email, book time, change price, move file, publish, delete.
• Reviewer: the person who owns the decision.
• Run receipt: what the AI did, which sources it used, and what changed after review.

This does two things at once. The team moves faster because recurring AI work does not have to be invented from scratch every time. The team also gets better control because nobody has to guess which chat contained the source material.

MCP connections need a tool list, not just a login

Notion describes MCP connections as a way for Custom Agents to read and act in external apps through Model Context Protocol. MCP, or Model Context Protocol, is an open way for AI tools to connect to data sources and actions. In Notion's help page, each connection can have tools switched on or off, and write tools such as create, update, delete, send and post require confirmation by default.

Source: MCP connections for Custom Agents

That is the right level for smaller teams. You do not need a ten-page security policy to start. You need a short tool list for each agent:

• which sources may the agent read?
• which actions may the agent suggest?
• which actions must never run automatically?
• who owns the account or permission used by the connection?
• where can someone read the run log if something goes wrong?

If the agent only summarizes customer questions, read access may be enough. If it creates HubSpot tickets, moves Box files or updates a Miro board, decide which write tools need a human yes. API keys belong in environment variables or a secret manager, not inside prompts or loose chat messages. Unneeded personal details can be redacted before they move further. That is not a brake. That is what makes the integration usable in real work.

Add a budget or run limit field to the board if you use Custom Agents. Notion's credits page says agents pause when credits run out and admins can receive warnings at 80 and 100 percent usage. For a smaller team, that is a useful signal: treat agent work as an operating cost, not magic.

Source: Notion credits

A 45-minute test: from meeting to visible AI work

Choose a workflow that already happens every week. Not the whole business. One meeting, customer request, quote follow-up or lesson plan.

Minutes 0-10: choose the board and statuses

Create or reuse a Notion database. Add Inbox, Ready for AI, Running, Needs review and Done. Decide that no AI output becomes external until the card passes Needs review.

Minutes 10-20: create the first card

Use a real but bounded example: notes from an internal meeting, a customer question you already answered, or a plan that does not contain sensitive details. Fill in desired output, sources, reviewer and stop rule.

Minutes 20-35: let AI produce a draft

Run the prompt below in Notion Agent, Claude, ChatGPT or the tool you already use. If you have External Agents in Notion, connect the run to the card. If not, paste the result back into the card. The point is visibility, not the button.

Minutes 35-45: review and write the run receipt

Answer three questions: What was useful? What needed editing? Which rule should be on the next card? If the answer is "AI guessed too much", that is not a failed test. It usually means the card needed better sources, a clearer stop rule or a narrower task.

Copy-paste prompt: turn a task into an agent card

Paste this into the AI tool you use. Replace the brackets.

You are my workflow assistant. Create an AI work card for our shared board.

Task: [describe the task]
Sources allowed: [links, files, meeting notes or database pages]
Audience or recipient: [customer, internal team, class, supplier]
Desired output: [reply draft, summary, checklist, follow-up email, quote material]

Return:
1. A short title for the work card.
2. What AI should do in no more than five bullets.
3. Which sources must be checked before the output is used.
4. Which actions require human approval.
5. Which details should be redacted or left out.
6. A first draft.
7. A run receipt: assumptions made, sources used, and what a human should review carefully.

Be practical. If the source material is too thin, ask for the missing details instead of guessing.

Three first boards for smaller teams

Customer questions to answer cards. Put recurring customer questions on the board. AI can read previous replies, terms and product notes. It can draft the answer, but a person approves before anything is sent.

Meetings to decision cards. Use AI Meeting Notes or ordinary meeting notes. AI can extract decisions, owners and next steps. Speaker labels help, but the reviewer still checks who promised what.

Course material to exercises. Let AI create a quiz, a simple HTML exercise or a checklist from an existing document. Publish only after the teacher or course owner has tested that the exercise makes sense without extra explanation.

What Hammer would do with this

For Hammer, Notion 3.6 is a reminder of a simple principle: start with the work surface before chasing more AI features. If AI work is visible on a board, each run can teach the team something. If it stays in private chats, improvement becomes personal, random and hard to scale.

In Mindset Forge, this is about choosing the right first workflow. In Tool Forge, it is about connecting Notion, Outlook, files, CRM or other tools with the right permissions. In Skill Forge, it is about making the routine repeatable: card template, prompt, stop rules and run receipt.

Start with one work card this week. If the board makes it easier to see what AI did, who reviewed it and what should happen next time, you have found a workflow worth building on.