Save AI-generated summaries to a review queue with Make

End the workflow at a review queue

The target is one candidate summary per source version, saved for review. Do not add publication destinations or mark generated text as published. Use a fictional memo: “Meeting: November 5 at 15:00. Proposed budget: ¥80,000. Approval is pending.” A reviewer must catch any output that changes a proposal into an approved decision.

Separate source, candidate and reviewed text

Prepare source_id, source_version and source_text in the input sheet. Create candidate_summary, review_status, reviewed_summary and reviewed_by in a review sheet alongside the source ID and version. Use the ID-version pair for processing identity so a source correction can become a new review item without overwriting an already approved summary.

Check content eligibility and input length

Remove unnecessary personal or confidential content before external processing and confirm that the source may be used for this purpose. Skip empty inputs. Send oversized material to a deliberate review path instead of silently truncating it: a missing final paragraph can remove a correction, exception or qualification that changes the meaning.

Configure Summarize text

Add Make AI Toolkit > Summarize text and map the approved source_text to Text to summarize. Select the connection and model, then inspect a controlled output before mapping it to storage. Do not connect modules that execute URLs or commands embedded in the source. Source instructions remain document content, not authorization to publish or send.

Save a candidate with a fixed review state

Use Google Sheets > Add a Row to store the candidate, source identity and fixed needs_review status. A model saying “verified” must not set an approval field. Add an accessible source reference for the reviewer. Set Value input option to Raw so a leading “=” or similar value is stored as supplied rather than interpreted as a formula.

Review facts and unresolved decisions

Compare dates, amounts, names, owners, negation and the distinction between proposed and agreed actions. In the example, retain the meeting time, proposed ¥80,000 budget and pending approval. Save edits in reviewed_summary, with reviewer and time, before changing the status to approved. Keep the original candidate for later quality analysis.

Treat long-document chunking as another design

Splitting material with Chunk text does not establish a correct whole-document summary. A larger workflow needs to check context across chunks, duplicated material and contradictory statements. Begin with short eligible sources and a clear length limit so the reviewer can actually complete the defined comparison.

Test version changes and replay

Exercise the same ID-version pair twice, a revised version, empty input, numeric details and unresolved or negated statements. Check candidate counts and state, and ensure regeneration cannot overwrite approved human edits. Review Make and provider charges and retain source identity when generation fails so the record can return to manual review.

• Route AI categories for human confirmation
• Manage source versions and replay
• Inspect storage after partial failure