How to use Notion AI: Notion Agent, Custom Agents and meeting notes, honestly explained, with 8 prompts to copy

Notion AI changed more in the last two months than in the year before. This guide explains what Notion Agent and Custom Agents actually do now, which plan you need, how the new credits work and the prompts that save real time.

The short answer

Notion AI is the AI built into Notion, the docs and database workspace made by Notion Labs. Its centre is Notion Agent, a chat assistant that reads and edits your pages and databases, plus Custom Agents that run on schedules and triggers. Since 19 October 2026 agents get their own computer and browser and can keep working on one task for a full day. The full features need the Business or Enterprise plan; Free and Plus only get a limited trial.

Made byNotion Labs, Inc., United States
Free planOnly a limited trial of Notion AI on Free and Plus
ModelsAuto router by default, or pick models such as Opus 5.5, Sonnet 5.5 and GPT-6 Luna
Best atAnswers from your own workspace, meeting notes, databases, recurring team reports
Shows sourcesMeeting summaries link back to the transcript; workspace answers draw on pages you can open
Team plansBusiness and Enterprise, with credits per seat and admin model controls
Your dataNot used for training by default on any plan. Enterprise gets zero data retention for most models.

What Notion AI is in November 2026

Notion AI used to mean a writing helper inside your pages. It is now a set of agents that work on your workspace. The main pieces are:

  • Notion Agent, the chat assistant you open from the small face icon at the bottom of Notion. It creates and edits pages and databases, answers questions from your workspace and connected apps, reads uploaded files and hands back spreadsheets or documents.
  • Custom Agents, background agents that run without you. They start on a schedule, when a database property changes, when a comment is added to a page, or when something happens in a Slack channel, such as a new message or an @-mention.
  • AI Meeting Notes, which transcribes and summarises meetings without a bot joining the call.
  • Research Mode, Autofill, AI blocks and Enterprise Search (Beta) for longer reports, filling database properties, writing inside pages and searching apps such as Slack, Google Drive and GitHub.

Three releases in a row reshaped all of this. Notion 3.7 on 15 September 2026 added skills, reusable instructions that teach the AI how your team works. Notion 3.8 on 13 October added connections to more than 150 tools through MCP, column permissions in databases and seven new models. Then on 19 October Notion rebuilt its agents: they now have their own computer and browser, a Review for me mode that checks actions before they run, and they can keep working on a single task for a full day (the release notes say up to 24 hours).

The same release introduced Auto, a router that picks a model for each task and switches if it gets stuck. Notion says that across more than 20 common tasks Auto gave results similar to Opus 5.5 at about 90% lower average cost. You can still pick a model yourself; the October picker includes Opus 5.5, Sonnet 5.5 and GPT-6 Luna, and Claude Fable 5 is available on Business and Enterprise once a workspace owner turns it on.

Which plan gets Notion AI, and how credits work

This is the part most people get wrong. On the Free and Plus plans, Notion AI is only a limited trial, with features like generating docs and autofilling databases. Notion Agent in full, Custom Agents, AI Meeting Notes and Enterprise Search come with Business and Enterprise.

All AI use now runs on one credit system. New customers got it on 19 October 2026 and existing customers move on 2 November 2026. Each Business and Enterprise seat includes 1,500 Standard credits and 500 Select credits a month:

  • Standard credits pay for work where Notion picks the model, which is what Auto does.
  • Select credits pay for work where you pick the model, and for Custom Agents.

The split is deliberate, so a busy agent cannot drain everyone's everyday allowance. Notion says most people will not run out. The ones who might are people who always choose a large model by hand, and teams running several Custom Agents. Admins can buy extra credits, set spending limits for members and agents, and get alerts as the workspace nears its limit.

One more option is new on Business and Enterprise: you can connect your ChatGPT Plus or Pro plan to your personal Notion Agent, so you are not paying twice for AI. It covers your personal agent, not Custom Agents. Check the Notion pricing page for current prices.

Where Notion AI is genuinely strong

Answering questions from your own workspace

This is the reason to use Notion AI instead of a general chatbot. Notion Agent sees what you can see in your workspace and nothing more, so "what did we decide about the pricing page in the last three weeks" gets an answer built from your actual meeting notes, docs and, if connected, Slack. A general assistant would need you to paste all of that in first.

Meetings that turn into work

Type /meet on a page in the desktop app and select Start transcribing. The desktop app records both your microphone and the call audio, which is why it is better than the browser for video calls. You get a transcript, a summary that links back to the transcript so you can check it, key points and action items, filed in a meetings database. It supports 16 languages including English, German, French and Spanish, with a limit of 10 hours per user per day. A Custom Agent that watches your meetings database can then file the action items in your task tracker.

Databases that fill and tidy themselves

Notion Agent can create databases, add views and properties such as relations, and query specific properties. Autofill fills properties across rows, such as a one-line summary or a category for every item in a feedback database.

Recurring team work with Custom Agents

Custom Agents are where Notion now differs most from other AI tools. Good first agents are a weekly status report that reads your project database every Friday, a triage agent that sorts new feedback or bugs, or a Q&A agent embedded on your team wiki. Since September a Custom Agent can hand parts of a job to sub-agents, each with its own instructions and model.

Where Notion AI falls short

  • No real AI on cheaper plans. If your team is on Plus, you only have a trial. That is the biggest catch.
  • There are things it cannot do. Notion's help centre lists them: Notion Agent cannot create or edit comments, share pages, change permissions, create database automations, templates or page layouts, or build formulas, rollups and buttons. You still do those yourself.
  • It acts as you. When it posts in Slack, the message comes from you. Review anything that goes to other people.
  • Credits are new. Until your team has a month of usage data, nobody knows how fast Select credits will go. Watch them in the first weeks of November.
  • Your workspace is the ceiling. If your pages are out of date or scattered, the answers will be too. Notion AI rewards a tidy workspace.

How to get better results from Notion Agent

  1. Point it at the right pages. Use @ to mention the exact pages, databases or people, and the All sources menu to limit where it looks. Narrow context gives sharper answers.
  2. Give it standing instructions. Click the agent's face in the bottom corner, click it again in the chat, and choose Add instructions. Put your role, your team and your preferred format there once.
  3. Turn good prompts into skills. When a prompt works, save it as a skill and run it later by typing / in the chat. Skills can be shared in a skills library and exported as SKILL.md files for other agents.
  4. Leave Auto on for everyday work. It uses Standard credits, and Notion says it gets results similar to Opus 5.5 on common tasks at a fraction of the cost. Pick a model by hand only for long, hard work.
  5. Use Review for me for anything that changes things. Notion describes it as a mode with safety checks, which matters once an agent can edit many pages at once.
  6. Start Custom Agents small. One agent, one trigger, a narrow set of pages. Notion itself suggests starting with one or two sub-agents at most.

Our Notion lessons on /learn take you through setting up instructions and your first skill in about ten minutes.

Notion AI privacy settings and where to find them

Notion says that by default your data is not used to train models, and its contracts bar its AI providers from training on it. Notion's security page lists no plan exceptions, so there is no training switch to hunt for. The difference between plans is retention: on Free, Plus and Business the AI provider keeps your data for 30 days or fewer before deleting it, while Enterprise uses zero data retention for most models, with the few models that keep data switched off until a workspace owner opts in.

The settings that matter live in Settings, under Notion AI: workspace owners turn on high credit consumption models such as Claude Fable 5 there, and AI connectors for Gmail and Google Calendar are set up and disconnected from the same Settings, Notion AI page. Since October, database owners can also set column permissions so that people and agents cannot see sensitive properties such as salaries. For AI Meeting Notes, always tell people they are being transcribed; Notion offers a consent message you can paste into the meeting chat.

Which Notion plan suits whom

  • Individuals on Free or Plus who mainly want writing help will find a general assistant such as ChatGPT or Claude gives more for less. Keep Notion for organising.
  • Teams already living in Notion should look hard at Business. Workspace answers, meeting notes and Custom Agents are worth it when your knowledge is already there.
  • Companies with sensitive data should choose Enterprise for zero data retention, HIPAA support and tighter admin controls.

Not sure Notion is the right tool at all? See Notion AI alternatives, compare it in Notion vs Gemini, or build a full set of tools on Stacks.

Prompts to copy

Catch up on a project

Read @[project page] and every meeting note from the last 14 days that mentions it. Tell me in under 150 words what was decided, what is blocked and who owns each open action. Link the page you took each point from.

You get: A sourced catch-up you can trust in two minutes

Meeting to tasks

From this meeting note, list every action item with an owner and a due date. Add each one as a row in @[tasks database], with Status set to Not started and a relation to @[project page]. Show me the list before you add anything.

You get: Action items filed in your tracker, after your check

Build a database from scratch

Create a database called Content calendar with properties for Title, Channel (select: LinkedIn, Newsletter, Blog), Owner (person), Publish date, Status and a relation to @[campaigns database]. Add a board view grouped by Status and a calendar view by Publish date.

You get: A working database with two views in one go

Draft a weekly status report

Using @[projects database], write this week's status update for leadership. One line per project with status, the change since last week and any risk. End with the three decisions we need from them. Keep it under 250 words.

You get: A leadership update built from live project data

Summarise customer feedback

Read the 50 newest rows in @[feedback database]. Group them into no more than six themes, give the count for each, quote one typical comment per theme and suggest one product change per theme.

You get: Themes, counts and examples instead of a pile of rows

Write a Custom Agent

Help me set up a Custom Agent. Every Friday at 4 pm it should read @[projects database], find projects whose due date has passed and are not Done, and post a short list with owners to [Slack channel]. Draft the instructions, trigger and the pages it needs access to.

You get: Ready instructions, trigger and access for your first agent

Prepare for a meeting

I meet [person or client] tomorrow about [topic]. Search my workspace and connected Slack for everything relevant from the last month. Give me five bullet points of context, two open questions and a suggested agenda.

You get: A one-page brief before the meeting

Turn a prompt into a skill

Turn this into a reusable skill called Exec summary: take any page, write a five-sentence summary for senior leaders that leads with the decision needed, then lists risks and next steps. Use British spelling and no jargon.

You get: A skill you can run with / on any page

Start the Notion lessons