Choose Basic Autofill when the answer is already in the database row or page. Choose Custom Agent Autofill when the job needs other workspace pages, web research, or several coordinated property updates. A meeting summary usually belongs in Basic. A project handoff checklist that depends on a separate policy page is a reason to consider Custom Agent.
I'd start with Basic for a small team's notes, briefs, and feedback. Its narrower scope makes the source easier to check, and Basic does not use Notion credits. Consider Custom Agent only when the task needs a wider source scope or coordinated actions.
Choose by the source and output
Notion's current Autofill documentation distinguishes page-level summaries, extraction, translation, and tagging from the broader search and multi-step work available through Custom Agent Autofill. Use that distinction to choose the mode:
| Your task | Start with | What the choice depends on |
|---|---|---|
| Summarize a meeting note or extract its next action | Basic | The necessary evidence is on that page. |
| Categorize feedback with select or multi-select options | Basic | The entry contains enough information to choose a category. |
| Find a project's requirements in a separate policy page | Custom Agent | Workspace search is enabled and the agent has access to the relevant sources. |
| Enrich an entry with current information from the web | Custom Agent | Web search is enabled; the result and its sources still need checking. |
| Write a summary and extract a due date in the same run | Custom Agent | The workflow needs coordinated updates to multiple properties. |
A longer prompt alone is a poor reason to upgrade the workflow. Ask two things instead: what information does the task need, and where does the result have to go? Those answers decide the mode.
A custom prompt does not mean Custom Agent
The naming is easy to misread. Inside Basic, Notion offers choices including Summary, Translate, Key info, and Custom autofill. That last option lets you specify your own page-based task. It does not turn Basic into a workspace-searching agent.
For example, “Extract the next action explicitly stated in this meeting note” is a custom instruction that can still fit Basic. “Find the client's handoff requirements in our policy database, then update this project” needs a source beyond the current page. Basic does not browse the web or search across all your Notion pages.
Set up Basic for a page-based field
Start with an output property you can review separately from the original notes. For a client meeting, a text property named Next action draft makes the distinction clear: the notes contain the evidence; the AI-filled value is a proposed extraction.
- Click the name of the property you want to fill.
- Select AI Autofill or Set up AI Autofill, then select Basic.
- Choose the task. Use Key info for extraction or Custom autofill when you need a particular output format.
- Choose when it should run and save. Notion documents manual runs, page creation, and page edits as Basic options.
I'd do a manual run first, before enabling a trigger that updates many records. That run shows whether the property answers the question you care about.
Consider this fictional meeting note: “Harbor Co. approved the draft. Next action: Maya sends the final brief.” A suitable instruction would be:
Extract only the next action explicitly stated on this page. Return one short sentence. If no next action is stated, return “Not stated.”
The expected extraction for that example is “Maya sends the final brief.” What you're checking is whether the AI-filled field keeps the stated action without inventing a deadline, recipient, or extra task.
If the page only says “follow up soon,” a confident-looking date is not an improvement. Keep missing information visible and resolve it with the person responsible.
When Custom Agent is worth the setup
Consider a project page with a Service tier value of “Standard.” Its handoff requirements live on a separate, approved policy page. Basic cannot retrieve that cross-page context. Custom Agent Autofill can use workspace search and richer instructions to look it up.
The setup begins from the same property menu: choose AI Autofill → Custom Agent. Write the job, name the intended sources and output properties, enable workspace or web search if needed, and choose the run timing. Notion's Custom Agent Autofill setup includes manual runs, page creation, page edits, and schedules.
For the handoff example, I would keep the instructions focused:
Use this project's Service tier and the approved Handoff policy page to identify the required handoff items. Write them in Handoff draft. If the tier is missing, the policy has no matching rule, or the sources conflict, write “Needs review” instead of inventing requirements. Only fill Handoff draft if it is empty.
Prepare the example before the agent fills anything: keep Service tier as the source field and Handoff draft as an empty text destination. On the approved policy page, use an explicit rule such as “Standard: final files and a handoff note.” In this example that's the entire rule; your own policy sets what your real service includes. The expected draft is “Final files and a handoff note,” with no added training session or deadline.
That prompt describes the behavior you want, but a run can still get it wrong. Check the draft against the matching policy before you treat it as the project's agreed handoff.
Check source access: linking a page in agent instructions does not grant access to it. Add the required pages or databases under Settings → Tools and access. Keep web access off for this internal policy lookup.
Multiple outputs are another valid reason to choose Custom Agent. It can, for example, summarize a brief and populate a Due Date property when the brief explicitly mentions a deadline. State what it should do when there's no deadline. Otherwise a field that looks complete can hide a guess.
Sometimes you do not need AI at all. If a project is already linked to a client and you only need to display a value from that client record, a relation and rollup can keep that value connected to its source. Reserve Autofill for work that requires interpreting content.
Review access before sharing the agent
Custom Agents have independent permissions, separate from individual users. Someone who can use an agent may receive information through it that they could not open directly. For a freelancer or small agency with several clients in one workspace, this matters more than the length of the prompt.
Give the agent the pages and databases needed for the job. Then check who can use it and who can read the destination property. Do not give a general client-facing agent access to private notes for every client just to make one lookup easier.
Notion says agents do not have full workspace access by default. Broad access is a configuration choice. A person with Can view and interact can inspect the agent's settings in read-only mode, including its instructions, resources, tools, and triggers. Use that visibility to review the actual configuration before relying on it.
Compare the cost of the workflow
Basic Autofill is included on Business and Enterprise plans and does not use Notion credits. Custom Agent Autofill uses credits based on the work needed to complete the task. Notion also requires Business or Enterprise for Custom Agents.
For a recurring workflow, consider how often it will run as well as what each run does. A one-time handoff draft and a scheduled enrichment job are different commitments. Start with the trigger you need, rather than automatically choosing every page edit or a frequent schedule.
Have the workspace admin check the current credit balance and how your billing cycle affects credits before enabling ongoing work. Custom Agents pause when the workspace has insufficient credits, until credits reset or an admin adds more. A blank field can therefore be a billing or run issue, rather than missing source information.
Check results and protect reviewed values
Do not make the first automated destination an already approved client commitment. Keep AI drafts separate while you decide whether the workflow is dependable enough for its role. Review clear entries, missing information, conflicting details, and a record with an existing value.
For Custom Agent Autofill, Notion documents the instruction “Only fill this property if it is empty.” Use it when you want to preserve populated cells, and confirm what happens to a human correction on the next run. The instruction describes how the agent should behave; access permissions remain a separate control.
| Prepared case | Required review outcome |
|---|---|
| Standard tier; one matching approved rule | Draft contains only final files and a handoff note. Reject extra promises not in the policy. |
| Service tier empty | Draft says “Needs review.” Ask the project owner for the tier instead of accepting an assumed package. |
| Policy missing a rule or offering conflicting rules | Draft says “Needs review.” Resolve the policy with its owner before treating any handoff as agreed. |
| Handoff draft already contains a reviewed value | The existing value remains as instructed. Compare it before and after the run; the blank-only instruction is not an access lock. |
Keep these cases in internal sample records, rather than using a live client promise as the first destination. A reviewer should know which rule supported the draft and who resolves “Needs review.” A filled cell without that evidence has not passed the check.
If Basic does not fill, check that AI Autofill is enabled for the property and that the page contains enough information. If it cannot find a fact from another page or the internet, change the mode and configure the necessary search and access; rephrasing a Basic prompt cannot expand its source scope.
For a failed Custom Agent run, a person with Full Access can inspect Activity for the trigger, actions, errors, and failures. Check the run timing, granted sources, instructions, and credits before widening permissions. When it does fill, compare the value with its source as well as the requested format.
Before turning on automatic runs, decide who checks the draft property and what happens to an entry marked “Needs review.” A field that prompts the right follow-up is more useful than one that merely stays full.
