A Custom Agent is not a chat window with a longer system prompt. It is a scheduled or event-driven process that reads your workspace databases, cross-references them with outside tools, and writes back. Notion connects them to Slack, Figma and GitHub through Model Context Protocol integrations. Typical jobs: flagging items that have sat in a backlog column too long, compiling a status report before stand-up, opening tickets from a meeting summary.
Notion owns the model routing. Admins on Business and Enterprise plans can specify which provider's model an agent is allowed to call and set a default. That control arrived in September 2026 and it matters if your data-protection team wants to know whether workspace contents are leaving for a US inference endpoint. The audit log records permission modifications, instruction edits and integration changes, which is the artefact a compliance reviewer will actually request.
Sub-agents in Notion 3.7
The 3.7 release lets a primary agent delegate a step to a specialized sub-agent. One practical shape: a triage agent receives an incoming bug report, hands log analysis to one helper and reproduction steps to another, then merges both results into a single ticket. Under the hood the pattern is the same multi-agent choreography OpenAI, Anthropic and Google have been pushing through their agent SDKs this year – a coordinator plus narrow specialists rather than one enormous prompt.
Meeting-note triggers landed in July 2026. A summarized AI meeting note can fire a Custom Agent that writes task tickets and posts a Slack recap without anyone typing a command. The full description of how these pieces fit together is in the UC Today report that prompted this piece.
What it costs
Custom Agents are gated behind Notion's Business plan. The source figure is roughly £240 per user per year, which converts to about €280 at recent rates. For a ten-person team, Notion's pricing page quotes £160 to £200 per month. Seat cost is the floor, not the total. Agent executions then consume credits at $10 per 1,000 runs, a rate introduced in May 2026. Sub-agent chains use more credits per completed task than a single-step prompt, because each delegated call is a separate run.
Put that next to raw model pricing and it looks unremarkable. GPT-6.1 Sol, Claude Sonnet 5.5 and Grok 4.7 all sit near $2 per million input tokens and between $6 and $10 per million output. DeepSeek-V4.1-Flash, open-weights and downloadable, charges $0.30 per million input at peak. What Notion sells is not the model. It is the fact that the model already knows where your sprint board lives and has write access to it.
European availability and the AI Act
Notion is a US company. Custom Agents process EU customer data on US-controlled infrastructure unless you negotiate a data-residency arrangement on Enterprise. For most European buyers that is a procurement conversation, not a toggle in settings. If your organisation has signed standard contractual clauses and run a transfer impact assessment for other US SaaS, Notion slots into the same paperwork. If not, this is the point where legal gets involved.
Since August 2026 the EU AI Act's transparency provisions are enforceable. Article 50 requires that people be informed when they interact with an AI system and that synthetic outputs be marked. A silent agent that updates a shared project database on a schedule sits awkwardly against that text, even if the update is technically logged. The Article 50 wording is short enough to read in ten minutes; the harder question is what a national supervisory authority will consider adequate marking for a changed database row.
The beta label matters more than it looks. Notion still carries it in its Help Centre, which gives the vendor room on reliability claims and gives buyers a reason to keep write access behind a review step. Read-only agents that only report are easy to justify. Agents that close tickets and post into shared Slack channels are a different approval.
What Rakuten France and Remote report
Two customer numbers are worth quoting. Rakuten France runs more than 80 Custom Agents, including four overnight backlog-check agents that post separate Slack reports for tech leads and PMs. The company says it has automated 58,000 workflows. Remote reports over 95 percent triage accuracy, more than a quarter of IT tickets resolved without a human, and 20 hours per week saved.
Both figures come from vendor customers describing their own deployments, so treat them as case studies rather than independent measurement – the same distinction we apply to every vendor claim at AI Arena. Neither company has published the cost side of those runs, and credit consumption is the number that decides whether an agent stays switched on.
How to start without burning credits
If you already pay for Notion Business, the agent builder is included and the only new line item is credits. Build one read-only agent first: a morning summary of overdue items, delivered to a channel you already read. Watch the audit log for a week. That gives you a real credit-per-useful-output baseline before anything is allowed to write into a database.
If you are on the Free or Plus tier, none of this is available at any price. Custom Agents are a Business and Enterprise feature, and sub-agent chains do not change that boundary.
Can Notion Custom Agents run without sending workspace data to US servers?
Only on Enterprise plans with a negotiated data-residency agreement. Business plans use Notion's default routing, which is US-operated.
Do Custom Agents replace Notion's built-in AI features?
No. Notion AI is a prompt-driven assistant that lives inside the editor. Custom Agents are separate processes triggered by schedules, events or meeting notes, and they bill per execution through the credit system.
How many credits does a sub-agent workflow typically consume?
Notion does not publish a per-workflow figure. Each delegated call counts as its own run, so a three-step chain costs more than a single prompt. The audit log shows what actually executed.
