Notion AI in databases: autofill and per-row summaries (2026)

If you already run a Notion database — CRM, backlog, candidates, content — Notion AI can work row by row: autofill properties, short summaries and prompts that respect the columns you pick. This is not the meeting-notes flow on a standalone page; it is making each record carry ready context without copying rows into another chat. This walkthrough takes about 15 minutes on a real base with at least five rows.

1. Steps 1

   Open the base you already use, not a demo copy

   Go to the base the team uses every week — table or board. Spot two or three columns that already hold operational truth (description, status, notes, link). This flow assumes the row is the unit of work; AI does not fix a chaotic schema.

   What you should see: Table or board view with real rows and columns with mixed but readable content.

2. Steps 2

   Add an AI (autofill) property with a short instruction

   From the column header, add an AI / Autofill property. Write a concrete instruction, for example: «In one sentence, what the client wants and the next step. Only use linked columns. If data is missing, write [MISSING].» Avoid «analyze this row» with no output shape.

   What you should see: A new AI column with the prompt saved and empty cells or a generate control on the first row.

3. Steps 3

   Choose which columns the AI may read

   In the AI property settings, select only the source columns you need — description, internal notes, URL, status. Do not tick the whole schema: every extra column is noise or PII you do not want in the summary. If attachments matter, reference them in a text column first.

   What you should see: A list of checked source properties (typically three to five), not the entire base.

4. Steps 4

   Generate a pilot row and compare it to sources

   On one representative row, run autofill in the AI cell. Read the output next to the source columns. If it invents a next step or a label that is not in the data, clear the cell, tighten the prompt or enforce [MISSING]. Do not batch until one pilot row is honest.

   What you should see: One AI cell with short text aligned with sources, or [MISSING] where data truly is absent.

5. Steps 5

   Add a second AI column for row summary

   Add another AI property, e.g. «Summary (1 line)», reading the same sources or the main long column. Prompt example: «Max 20 words. Fact + blocker. No empty adjectives.» Useful for boards and views where nobody opens the full description.

   What you should see: Two complementary AI columns: one operational (next step) and one for quick scanning (one line).

6. Steps 6

   Work from a filtered view with the same prompt

   Duplicate the view or create a filtered one — e.g. «Status = In review» or «Summary empty». From that view, fill AI cells one by one or on a selection. The filter does not change the prompt, but it focuses work: only the rows that matter today. Note in the view description which AI column gets reviewed each week.

   What you should see: A view with an active filter, a subset of rows and AI columns ready to generate without endless scroll.

7. Steps 7

   Fill the empty batch and mark what you reviewed

   Filter rows where the AI column is empty and status is not «Archived». Run autofill on a small batch (5–15 rows). Walk each result: clear, edit manually or keep. Add an «AI reviewed» checkbox or move status to «Ready» only after your pass. AI proposes; you certify.

   What you should see: Most AI cells filled; checkbox or status updated on validated rows.

8. Steps 8

   Freeze the schema and document the prompt on the base

   When the prompt works, paste the final instruction in the base description or a linked «How to read AI columns» page. Lock schema editing if your plan allows, or agree not to change prompts without notice. New rows should follow the same ritual: sources first, autofill second, human review.

   What you should see: Base description or wiki page with the prompt text and which columns are source of truth.