David Neira is an official Notion Ambassador and the founder of NotioStore, where he builds custom Notion operating systems for architecture, construction, and interior design firms. He run 40 Notion Setup Sessions per week and writes here about making Notion hold up under real client work.
Every week, someone on your team does the same thing: reads through a set of records, pulls out the relevant information, drafts a summary, and either updates a database or sends a communication. It takes 30 to 60 minutes. It requires attention but not judgment. It happens every week without fail because if it doesn’t, something downstream breaks.
That task is an agent.
Notion AI Custom Agents are autonomous workflows that run inside your Notion workspace. You describe the job, set a schedule or trigger, give the agent access to the databases and tools it needs, and it executes. Every week, every day, or the moment a specific event occurs without you being at your desk.
This article covers what Custom Agents actually are, when they’re worth building, and how to build one that works reliably.
Before building anything, understand the distinction between Notion’s two agent interfaces because most of the confusion about “Notion AI Agents” comes from conflating them.
The personal Notion Agent: the AI assistant accessible from the bubble at the bottom of your screen in Notion. It’s interactive: you type a prompt, it does something, the session ends. Useful for on-demand tasks, querying your workspace, updating a record, drafting a page from context. It only works when you’re there to talk to it.
Custom Agents: proactive, autonomous AI teammates that run on schedules, triggers, and external events. You build them, configure them, and they run whether you’re there or not. A Custom Agent reads your daily logs every Monday morning and drafts a weekly recap. A Custom Agent monitors your Leads database and sends a Slack notification when a high-priority lead hasn’t been contacted in 5 days. A Custom Agent reads incoming Notion Mail messages and creates structured intake records in a database.
This article is about Custom Agents, the proactive, buildable layer. Custom Agents are available on Business and Enterprise plans and run on Notion credits, an add-on priced at $10 per 1,000 credits starting May 2026.
The most common mistake when building Custom Agents is scope creep at the design stage. An agent that “monitors all our databases, summarises activity, updates client records, and sends Slack updates” is an agent that will fail unreliably across several different tasks simultaneously.
The principle is simple: one agent, one job. If you can’t name the job in 3-4 words, the agent is scoped too broadly. “Weekly Field Ops Recap,” “New Lead Notifier,” “Post-Meeting Task Creator.” Each name describes one specific, bounded workflow. When the agent fails or produces unexpected output, a single-job scope makes debugging straightforward. When it works, its output is predictable and its value is measurable.
The prerequisite that applies before any of this: Custom Agents are the final layer of a mature Notion setup, they require a solid data foundation to deliver real value. An agent reading a messy, inconsistently named database returns inconsistent output. The value of an agent scales with the quality of the data it reads. Build your databases first, structure them well, populate them consistently. Then add agents.
Custom Agents run on one of four trigger types:
Schedule: runs at a defined interval: every day at 8am, every Monday morning, the first of every month. The most common trigger for recurring operational tasks. Your weekly summary agent runs on a schedule.
Database event: runs when a record in a specific Notion database changes. When a new record is created, when a property changes to a specific value, when a checkbox is checked. Useful for workflow automation triggered by team actions in the database.
Calendar event: runs when something happens in your calendar (an event is created, updated, or deleted). Useful for automations tied to meetings and scheduling, such as creating a meeting prep page when a new client call is booked, or logging follow-ups when an event changes.
Slack message: runs when a specific keyword or phrase appears in a connected Slack channel or DM. Useful for intake workflows: “when a support question appears in #client-questions, create a structured record in the Support Requests database.”
Mail: runs when an email arrives in Mail matching defined criteria.
For most small business applications, Schedule and Database event cover the majority of useful agent workflows.
Step 1: Go to the Agents section in the sidebar.
In your Notion workspace, find “Agents” in the left sidebar. Click the + button to create a new agent. You’ll be offered three starting options: describe what you want in plain language (Notion AI drafts the agent for you), start from a template, or start blank.
For a first agent, use the plain language description. Describe the job in one or two sentences: “Every Monday at 8am, read all Daily Log records from the past 7 days in the Jobs database, identify any issues flagged, and create a Weekly Recap page summarising what each active job accomplished and what issues need follow-up.”
Notion AI drafts a set of instructions, suggests a trigger, and proposes which databases to connect. Review all three, the draft is a starting point, not a finished agent.
Step 2: Name the agent for the job it does.
Name = the job. Not “My AI Assistant.” Try “Weekly Field Ops Recap” or “Lead Follow-Up Notifier.” The name appears in the agent directory and in trigger logs, when your team sees it, they should immediately know which workflow ran.
Step 3: Write precise instructions.
The instructions field is the most important part of the agent. Vague instructions produce vague output. The structure that works:
The more specific your instructions, the more predictable and useful the output. Treat the instructions field like a job description for a new team member who will follow them literally.
Step 4: Set the trigger.
Choose the trigger type and configure it. For a schedule trigger: select the days and time. For a database event trigger: select the database, the event type (record added, property changed), and the specific condition (Status changed to “Complete”). For a Slack trigger: connect the Slack workspace, select the channel, and define the keyword condition.
Step 5: Set the access scope.
Define exactly which databases, pages, and tools the agent can read and write to. Narrow access is better than broad access: an agent that only needs to read the Daily Logs database and write to the Weekly Recaps database should have access to exactly those two databases and nothing else. Scope permissions precisely, giving each agent access only to the databases, pages, and tools it needs.
If the agent needs to post to Slack, connect the Slack integration in this step. If it needs to read from Google Drive or interact with Linear or HubSpot, connect those tools here.
Step 6: Test before enabling.
Run the agent manually using the test function before turning on the trigger. Review the output against your instructions. Common issues on first run: the agent read more records than intended (tighten the filter conditions in the instructions), the output format is inconsistent (add more specific format instructions), or the agent tried to write to a database it didn’t need (tighten the access scope). Iterate on the instructions until the test output is exactly what you’d want a team member to produce.
Step 7: Enable and monitor.
Turn on the trigger. For the first two weeks, review every output the agent produces. Check whether it’s reading the right records, producing the right structure, and handling edge cases correctly. Adjust instructions as patterns emerge. After two weeks of clean output, reduce to spot-checking.
Every run is logged, so changes are visible and reversible. If an agent produces incorrect output, you can see exactly what it did and adjust the instructions before the next run.
1. Weekly Operations Recap
Job: Every Monday morning, read the past week’s activity across key databases and produce a structured summary.
Trigger: Schedule – every Monday at 7:30am.
Reads: Tasks completed in the past 7 days, meetings held in the past 7 days, any new client records created.
Produces: A new page in a Weekly Recaps database with sections: Work Completed, Client Activity, Open Issues, and This Week’s Priorities (pulled from tasks due in the next 7 days).
Value: The Monday morning briefing that currently takes someone 45 minutes to compile happens automatically before anyone sits down.
2. New Lead Notifier
Job: When a new record is added to the Leads database, send a Slack notification to the assigned salesperson.
Trigger: Database event – new record added to Leads.
Reads: The new Lead record (Name, Company, Source, Assigned To).
Produces: A Slack message to the assigned person: “New lead: [Name] from [Company] via [Source]. Open their record: [link].”
Value: Leads get acknowledged within minutes of entering the system, not the next time the salesperson happens to open Notion.
3. Post-Meeting Action Creator
Job: When a Meeting record’s “Summary Complete” checkbox is checked, read the meeting summary and create individual Task records for each action item.
Trigger: Database event – “Summary Complete” checkbox is checked on a Meetings record.
Reads: The meeting’s summary page, specifically the Action Items section.
Produces: New records in the Tasks database, one per action item, each linked to the meeting record, with the action description and the assigned person pre-populated.
Value: The action items from a meeting exist as actual tasks in the system the moment the summary is marked complete, not when someone manually creates them later.
4. Stale Lead Alert
Job: Every weekday morning, check the Leads database for any leads with Status not Won/Lost that haven’t been contacted in 5 or more days, and send a Slack alert to the assigned salesperson.
Trigger: Schedule – every weekday at 8:00am.
Reads: Leads database, filtered to Status ≠ Won/Lost and Last Contact Date ≤ 5 days ago.
Produces: A Slack message listing the stale leads with the assigned person mentioned: “@L.Torres – 3 leads haven’t been contacted in 5+ days: [names and links].”
Value: The follow-up queue that currently requires someone to check a filtered view every morning runs automatically.
Tasks requiring judgment about people or relationships. An agent can identify that a lead hasn’t been contacted. It cannot judge whether the lack of contact is intentional, perhaps you’re giving that prospect space after a difficult conversation. It cannot read the subtext of a client relationship. Any task where the “right” action depends on context that isn’t in the database is not an agent task.
Tasks with unpredictable input formats. If the data an agent reads is inconsistently structured: some records filled out fully, others half-complete, property names used differently across the team, the agent’s output will be inconsistently useful. The quality of an agent’s output is bounded by the quality of the data it reads.
High-stakes actions without human review. An agent can draft a client email. It should not send that email without a human reviewing it. An agent can create a follow-up task. It should not send a proposal. Build agents for the drafting and flagging steps; keep humans in the loop for the sending and deciding steps.
Tasks that require real-time external data. Custom Agents run on schedules and triggers, not in real-time against live external feeds. An agent can read your databases as they exist when it runs. It cannot monitor a live stock price, pull current weather data, or react in real-time to an external API event without an integration specifically built for that purpose.
Custom Agents run on Notion credits: $10 per 1,000 credits starting May 2026. Each agent run consumes credits proportional to the complexity of the task and the volume of data read and written.
Notion provides a credits dashboard showing usage by agent. Set per-agent credit limits with agent creators to keep spend predictable. If an agent is pacing higher than expected, you can disable it right away in Settings → Notion AI → Agents, then work with the creator to adjust it before turning it back on.
The practical guidance: start with low-frequency agents (weekly schedules rather than hourly) and lower data volumes (reading 10 records rather than 1,000) until you have a sense of the credit consumption for your specific workflows. Expand frequency and scope once you’ve validated the output quality and understand the cost.
For most small businesses running 3–5 practical agents on weekly or daily schedules, the credit cost is modest, typically well under $50/month. For teams running high-frequency agents reading large databases, the cost scales accordingly.
The best way to understand what Custom Agents can do for your specific operation is to build one for the most time-consuming recurring task in your workflow.
Identify the task: what does someone on your team do every week that requires reading existing information and producing a structured output? That’s your first agent. Build it, test it, run it for two weeks, and evaluate whether the output is reliable enough to trust.
One working agent that saves 45 minutes per week pays for the Business plan’s AI investment. Two or three reliable agents running in parallel change how your team relates to operational overhead.
The businesses that will get the most from Notion AI Agents are not the ones with the most technical skill. They’re the ones with the clearest picture of which recurring tasks don’t require human judgment and the discipline to build the database structure that gives agents useful material to work with.
A pile of tasks isn’t task management. Real task management is knowing, at any moment, what to do next, and that takes more than a list. This is a step-by-step tutorial for using Notion for task management.
Entrepreneurs can replace expensive tools with Notion by consolidating the overlapping ones, notes, docs, projects, wiki, simple CRM, into a single subscription, while keeping only the specialized tools that genuinely earn their price.
New to Notion? Discover the most common beginner mistakes and how to avoid them so you can build a smarter, simpler, and stress-free Notion system.
Built for the businesses that build everything else.
No product pitches. No padding. One weekly read for serious operators.