Notion AI · Practical walkthrough

Fill a Notion database property with AI, then audit what it wrote

You have a database of 180 supplier pages, each one a wall of notes, and a column called Summary that nobody has filled in since March. Notion AI can fill it. The question worth answering first is how you'll know whether what it wrote is true.

Task by Task14 min read + practice

What you'll make

one AI-filled property on a small practice database, with each sample value checked against its source page and a visible record of who checked it.

What you'll need

a Notion workspace on the Business or Enterprise plan (Basic Autofill is included and doesn’t use Notion credits), and a database whose pages have real content in the page body. About 30 minutes. Keep employee matters, health, pay and anything similar out of this: an AI-filled column gets copied, exported and shared without much ceremony.

Example result.

Northfield Packaging · Renewal 1 March 2027, stated in the signed terms.
Garrick Logistics · Suggested only, not agreed.
Brampton Print · No date stated on page.
Checked by: Sam · Checked on: 10 Oct · Source version: page edited 8 Oct

1. Choose one property and decide what “correct” means

Don’t start by filling five columns. Start with one, where a wrong value would be noticed rather than inherited, and write a sentence describing a correct cell. “A two-sentence summary of what this supplier provides and the terms stated” is checkable. “A summary” isn’t, and auditing against your own vague expectations is how a column of fluent nonsense gets approved.

Then copy five representative pages into a small practice database with the same properties, leaving the original untouched. Include an ordinary page, a missing field, a proposed date, a correction and a long page. Write the expected answer for each by hand. That’s your answer key.

Why this works: an answer key written before the AI runs is the only check that doesn’t share the AI’s assumptions.

What you should see: a five-page practice database and five expected answers.

2. Set up Basic Autofill

Notion has two kinds of autofill. Basic reads only the row or page it’s filling, with no web search and no other pages, which makes its answers easy to check. Custom Agent can search the workspace or the web and uses credits. Start with Basic.

In the practice database:

Hover over the property and click its name.
Select AI Autofill (or Set up AI Autofill), then Basic.
Choose what it should do: Summary, Translate, Key info or Custom autofill.
Choose when it runs: manually, on page create or on page edits. Pick manual for now, then save.

For Custom autofill, give it an instruction, not a topic:

Extract only dates that the page states were agreed, scheduled or contractually set, with what they're for. Do not extract dates that are suggested, proposed or discussed without agreement. If a date appears only as a suggestion, write "Suggested only, not agreed". If there is no date, write "No date stated on page". Use only what is written on this page.

Why this works: a property asked for dates will produce dates, and the easiest way when the page is silent is to write something true-sounding. A fixed phrase for each gap makes absence visible, and you can filter for it.

What you should see: the property set to Basic, manual runs and your instruction saved.

3. Run it on the sample and compare cell by cell

Run it on the five practice pages only. Check the scope before confirming. Then compare each cell with your answer key and the page, looking for:

Wrong detail: a date, figure or name that isn’t on the page, or belongs to something else.
A gap filled confidently: the page is silent, and the cell isn’t.
Right facts, wrong emphasis: on a page where the important sentence is the last one, such as “we agreed to exit this contract”, a summary of the first half could change someone’s next step.
Silence: an empty cell. Notion’s advice is to check Autofill is on for the property, then that the page has enough information. Empty isn’t proof the page has nothing relevant.

Why this works: each kind of error needs a different fix, and you can only choose the fix once you know which it is.

What you should see: five cells, each marked as matching the answer key or with a named error.

4. Keep the audit visible

Add ordinary properties for Checked by, Checked on and Source version. Mark a value reviewed only after a person has compared it with the page. If the page or the AI value changes, clear the check until someone reviews it again. A checkbox doesn’t reset itself because the meaning changed.

Use the column to help people scan, not to drive who gets contacted, paid or escalated unless every row has been read. And if the dates already live in a contracts system, link to that system instead of keeping AI-inferred copies of it.

Why this works: a column that runs on page edits can change after you checked it. The record shows which version you checked.

What you should see: the checked rows carrying a name, a date and a source version.

Check the result

Every sample cell matches your answer key or has been corrected.
Gaps show your fixed phrase, not an invented value.
Suggested or proposed dates are labelled as such, not as scheduled.
Each checked row records who checked it and which version of the page.

The usual failure: a suggestion becomes a fact. The month comes from the page, but the agreement doesn’t, and the cell reads “Review scheduled for November 2026”. Spot it by opening the page for every date and reading the whole sentence it came from.

For this page, the cell says "[cell value]". The page says "[source sentence]". Fill the property again using only what the page states was agreed. If the date is only suggested, write "Suggested only, not agreed".

[cell value] is what Autofill wrote. [source sentence] is the sentence on the page. Use this as a one-off Notion AI request on that page, then tighten the property instruction and rerun the sample.

Worked example: a review date that was never agreed

Fictional example. A Suppliers database has a Key dates property on Basic Autofill. The three pages say:

Northfield Packaging: signed terms section states renewal on 1 March 2027.
Garrick Logistics: "Priya suggested we review this in November, but we didn't agree a date."
Brampton Print: describes print services; no review or renewal date given.

With a first, looser instruction (“any renewal or review date on the page”), Autofill wrote: Northfield, renewal 1 March 2027. Garrick, review scheduled for November 2026. Brampton, no date stated on page.

The failure. Garrick’s page has a suggestion, no agreement and no year. The cell invented both the scheduling and 2026.
The repair. With the step 2 instruction, Garrick came back as “Suggested only, not agreed”. Expect improvement rather than a cure: a one-line cell has no room to show its reasoning, so keep a short supporting quote beside any date that matters.
The useful gap. Brampton’s “No date stated on page” is right, and filtering for that phrase lists every supplier missing a recorded date. The date may exist elsewhere, so check before concluding anything about the contract.

Task kit for this job: Download (ZIP) · Reusable instructions

Go deeper

Optional training: when Basic isn't enough

Custom Agent Autofill handles richer instructions: workspace search for related context, web search with sources when enabled, conditions such as “Only fill this property if Status is In Progress”, and several properties in one run. Set it up the same way but choose Custom Agent, enable search in the agent settings if you need it, and add “Only fill this property if it is empty” to leave checked cells alone. It uses Notion credits. If it can search, a cell may draw on pages other than its own, so build a new answer key from the same sources and audit a larger sample.

Optional training: a ten-row re-audit every quarter

A column that was accurate in October isn’t self-maintaining, especially if it runs on page edits. Pick ten rows at random (not the ten you remember, not the ten at the top), read the pages and compare. Record the result where people can find it, for example: “October: 10 rows checked, 1 wrong (Garrick: suggested date recorded as scheduled). Instruction tightened.” That gives you something concrete to say when someone asks whether the column can be trusted. Ten rows can reveal a problem, but they don’t estimate the error rate. Asking the AI to check itself may find candidates for review. A second draft is not a second witness.

Optional training: build the database with AI first

If the database doesn’t exist yet, create a new page and choose Table under Get started with, or type /database on a page, then select Build with AI. Describe it, review the preview, select Continue, choose a view and select Done. It’s free on Business and Enterprise. It can create new databases but can’t edit existing ones, or create their pages, automations, forms, charts or page templates. Let it draft a first structure, then take over.