Stop moving meeting decisions by hand: build a post-meeting agent in Notion

A meeting can end on time and still consume the rest of the day. Someone has to move decisions into the project plan, write a recap for Slack, create tasks, and check who promised what. It is an odd gap: AI can already write the notes, yet a person still becomes the conveyor belt between the notes and the work.

Notion's July 31 update closes that gap. A Custom Agent can now start when an AI Meeting Note has been summarized. The trigger is called Meeting note summarized. The agent can then update a project overview, draft an internal recap, or create tickets based on what people actually said.

Source: AI Meeting Notes can now trigger Custom Agents, Notion, July 31, 2026.

This is more useful than another meeting summary. For a smaller team, the post-meeting handoff is often where work gets lost. Start with one recurring meeting type and give the agent a clear contract for what it may read, what it may write, and what it must leave for review.

What a post-meeting agent actually does

A Notion Custom Agent is an AI workflow that runs in the background after a trigger or on a schedule. It uses the Notion pages, databases, and connected tools that you explicitly allow it to access. That makes it different from Notion Agent, which you start for an individual task while you are working.

The new chain works like this:

1. Notion AI transcribes the meeting and creates a summary.
2. The Meeting note summarized trigger fires when the summary is ready.
3. The Custom Agent reads the meeting note and follows its fixed instructions.
4. The agent writes to selected databases or channels and leaves a run record in Activity.

Notion's own examples include updating project status, posting a recap to Slack or email, and turning feedback into tickets. Custom Agents can also react to database changes, comments, schedules, and Slack events. They are available on Business and Enterprise plans and use Notion credits.

Source: Custom Agents in Notion, Notion Help Center.

The summary does not have to be a loose block of prose. AI Meeting Notes supports custom summary instructions for meeting types such as sales calls, stand-ups, and team meetings. The summary also contains citations back to the transcript. That gives the agent better evidence than a vague prompt that simply says "create tasks."

Start with the handoff, not the technology

Pick a meeting where the same follow-up happens every time. A weekly delivery check-in is usually a better candidate than an open strategy discussion. A customer call can work too, but have the agent create drafts and internal tasks before it sends anything outside the company.

Write down the handoff you want to exist when the meeting ends. It should answer five questions:

• Which meetings may start the workflow?
• Which fields should the agent create or update?
• When may the agent write directly?
• When does a person need to approve the result?
• What should a useful run note contain?

For a project check-in, the output might be an updated project record with status, decisions, blockers, and the next review. A sales call might produce a draft account note plus tasks with explicitly stated owners and dates. A school team could use an action register from a staff meeting, including questions that still lack a decision.

Do not let the agent fill empty fields with likely answers. If nobody named an owner, the field should say Owner missing. If no date was agreed, it should say No date agreed. The automation then becomes an honest handoff instead of a machine for polished guesses.

A 45-minute test for one recurring meeting type

This test requires Notion Business or Enterprise, AI Meeting Notes, and access to Custom Agents. Run it on a computer. AI Meeting Notes supports Swedish and several other languages, although speaker labeling is currently available only in English.

Minutes 0–10: choose the meeting and the outcome

Choose one recurring meeting type. Write one sentence that describes what should exist five minutes after the meeting, for example:

The project record should contain today's decisions, open blockers, and new tasks with an explicitly named owner and deadline.

Choose a destination page or database. Create the required fields before you build the agent. For tasks, title, source, owner, deadline, status, and review requirement are often enough.

Minutes 10–20: make the meeting note usable

Create or select a database for AI Meeting Notes. Add custom summary instructions that ask Notion to separate decisions, actions, blockers, and unanswered questions. Do not ask the summary to interpret tone or infer responsibility.

Notion requires participants to consent before a meeting is recorded and transcribed. The service provides text and voice options for the consent message. AI Meeting Notes supports Swedish. The feature remains in beta, and both AI Meeting Notes and Custom Agents require Business or Enterprise.

Source: Take AI meeting notes in Notion, Notion Help Center.

Minutes 20–35: build the agent

Open Agents, create a Custom Agent, and add the instructions below. In Settings, choose Triggers, add a Notion trigger, and select the event for a completed AI Meeting Note. Limit it to the meeting type or database used in this test.

Give the agent access to the meeting database and the specific project or task database it needs to update. If it should write to Slack, an administrator connects Slack and selects the relevant channels. The agent does not need access to the whole workspace to move one decision to the right place.

Minutes 35–45: run and review

Use a real internal check-in or an earlier audio recording. Notion can accept AAC, M4A, MP3, and WAV in a Meeting Notes block. The agent should start automatically when the summary is ready.

Open Activity. Check what triggered the run, which actions the agent took, and whether any errors occurred. Compare the output with the transcript citations. Adjust the instructions before you add more meeting types or enable external messages.

Copy-ready instructions for the agent

Paste this into a new Custom Agent's instructions and replace the bracketed fields.

You are the post-meeting operator for [MEETING TYPE]. You start when an AI Meeting Note has been summarized.

Your job:
1. Read the summary, handwritten notes, and available transcript citations.
2. Use facts from this meeting only. Do not fill missing facts with assumptions.
3. Update [DESTINATION DATABASE OR PROJECT PAGE] with:
   - decisions that were actually made
   - new actions
   - blockers
   - unanswered questions
4. For every action, record the task, explicitly named owner, explicitly agreed deadline, and link to the meeting note.
5. If an owner or deadline is missing, write "Owner missing" or "No date agreed". Do not guess.
6. Do not merge two actions if they have different owners, dates, or intended outcomes.
7. Create a draft internal recap in [CHANNEL OR PAGE]. Do not send anything externally.
8. Put anything that changes a customer promise, price, scope, or external deadline under "Approval required".
9. Finish with a short run note: source meeting, records created or changed, missing facts, and what a person needs to review.

If the evidence contradicts itself, do not change the destination record. Describe the conflict under "Approval required" and link to the relevant parts of the meeting note.

These instructions cover two details that generic agent prompts often miss. They define a concrete output format and tell the agent what to do when the evidence is incomplete. The second detail matters. An empty deadline field can be resolved. An invented deadline can turn into a customer problem.

Three workflows worth trying

The delivery meeting that updates the project

Have the agent move decisions, blockers, and actions into the main project record. The team reviews only the Approval required section. This reduces copying without hiding uncertainty.

The customer call that creates an internal draft

Have the agent write an internal account note and create tasks. Keep the external follow-up as a draft until the owner has checked price, scope, and dates. This integrates the agent into CRM work without giving it control over customer commitments.

The school team meeting that keeps unanswered questions visible

Have the agent collect decisions and actions, then create a separate list for questions that need a principal's decision, more evidence, or contact with a guardian. Missing decisions stay visible instead of disappearing inside a long summary.

Measure one week before you expand it

Run the workflow on the same meeting type for one week. Record how many meetings triggered the agent, how many tasks it created correctly, how often owners or dates were missing, and how many manual corrections were required. Check whether the team actually uses the run note.

Custom Agents have their own permissions. Someone using the agent may receive information from resources the agent can access even if that person cannot open the source page directly. Select specific pages, databases, and Slack channels. Check Tools and Access, share the agent with the right people, and use Activity as a normal operating check. The agent can then work on real projects without receiving broader access than the job requires.

Source: Custom Agents sharing and permissions, Notion Help Center.

If the test works, the next step may be to connect the reviewed handoff to a project tool, CRM, or internal report. That is where Tool Forge fits: clear fields, scoped permissions, approval steps, and a log that shows what the agent did. A successful run leaves the project record updated before anyone opens another tab.