walkthrough · sales

Build a meeting-to-follow-up draft workflow in Make

Build an illustrative Make meeting-draft design with synthetic notes, structured commitments, evidence checks, duplicate handling, and no automatic sending.

Author: Founder identity pending confirmation
Published: Pending first public release · Fact-check: Human review pending
intermediate · build · Approximately 5 minutes to read
Evidence: illustrative · Maintenance: needs retest
Action authority: draft · Data: public
Approach: AI-assisted workflow
Last tested: Not yet tested

Outcome and scope

Read synthetic meeting notes from Google Sheets in Make, extract supported commitments, and write a draft to a separate review sheet. The sequence is fixed. There is no automatic sending step and no email connection is needed for this design.

Evidence: illustrative. Verify against the live tool before acting. These are configuration steps and expected results, not a tested implementation. No screenshots, model settings, successful test date, plan entitlement, or setup-time claim is supplied. Do not connect live meeting records based on this page.

Stack and permissions to verify

Use Make’s hosted environment, a dedicated Google account or appropriately isolated test sheets, and an approved OpenAI API connection. Read only test notes, send only necessary synthetic text to the model, and write only the review sheet. Verify actual connection scopes, current operations, model configuration, retention settings, and charges before use.

Make’s Sheets documentation describes the integration. Its availability does not prove your chosen permissions or this design’s behavior. Stop if requested access is broader than the process owner permits.

Preparation

Read the meeting playbook. Make input columns meeting_id, notes_version, notes, and source_owner. Make review columns review_key, commitments, unknowns, draft, evidence, status, and reviewer_note. Keep input and output separate so an output cannot trigger itself.

Use meeting_id:notes_version as the review identity. Keep a ledger of pending, completed, and uncertain writes. For initial tests, run one item at a time under supervision; a row search does not prove safe concurrent uniqueness.

Sample input and output

Synthetic input M17, version 1: “Alex will send the outline. Priya will check availability for Friday. No deadline agreed.”

Expected structured target, not an observed model response:

{"commitments":[{"owner":"Alex","action":"send the outline","deadline":null,"evidence":"Alex will send the outline."},{"owner":"Priya","action":"check availability for Friday","deadline":null,"evidence":"Priya will check availability for Friday."}],"unknowns":["Delivery deadline not agreed"],"draft":"Alex will send the outline. Priya will check Friday availability."}

A person must resolve any relevant date or recipient ambiguity before sending. The application sets status to pending outside model output.

Step 1: select a single test input

Create a Make scenario with scheduling disabled. Configure a current supported Sheets read operation to select M17 from the dedicated sheet. Inspect the returned fields before adding the model step. Do not depend on an unverified new-row trigger to notice later edits.

Checkpoint: Only M17 version 1 appears. Missing notes, ID, or version goes to an exception state.

Step 2: establish the review key

Build the key from validated source fields. If the same key is complete, do not append another draft. If an earlier write is uncertain, reconcile the review sheet first. A changed notes version creates a new review requirement and marks the earlier draft as superseded.

Checkpoint: Sequential replay creates no additional logical draft. Concurrent delivery remains a separate acceptance test requiring verified duplicate controls.

Step 3: request extraction and a draft

Supply instructions separately from the notes. Ask for action, owner, deadline or null, evidence text, unknowns, and a draft using only supported commitments. Explicitly distinguish requests from agreements. Do not let the model set recipients, approval, credentials, or tool actions.

Checkpoint: “Friday availability” does not become a confirmed Friday start. “No deadline agreed” produces a missing deadline, not tomorrow’s date.

Step 4: validate and store safely

Validate the structured response and required fields. Reject invalid structure and unsupported evidence. Copy the meeting ID and version from the original input, not model-generated values. Store text as text so content beginning with formula characters is not executed by the sheet.

Write the draft and source references to the review table, then record completion in the ledger. If the write may have succeeded but its response was lost, mark it uncertain and inspect the destination before replaying.

Checkpoint: The row says pending and retains unknowns. It contains neither an approval decision from the model nor a claimed sent status.

Step 5: perform human review

Compare every commitment against notes using the follow-up review guide. Accept, edit, or reject. Verify the intended recipient and attachments in the normal sending process. A later draft edit requires review of the changed version.

Checkpoint: Rejecting or ignoring a draft causes no external action. The scenario has no email-sending step.

Normal and edge-case tests

Record observed outputs and human verdicts separately from targets. Add a changed-notes test and inspect all statements, not just the overall tone.

Troubleshooting and stop/recovery

If rows repeat, inspect keys, trigger selection, and incomplete writes before rerunning. If dates appear without evidence, hold affected drafts, correct instructions or source preparation, and add a regression case. Consult incomplete executions for current platform behavior; verify settings rather than assuming failures are retained automatically.

Stop scheduling and return to manual follow-up during repair. Reconcile existing drafts, preserve keys, and rerun affected tests before resuming.

Completed-work checklist and maintenance

Confirm isolated sheets, stable keys, supported commitments, visible unknowns, safe text writes, pending review, no send step, and tested manual fallback. Record actual account and model configuration during reproduction. Review corrections, duplicate states, access, and usage; rerun fixtures after source, prompt, or connector changes.

Sources and next step

Official Make references above support integration verification only. This workflow remains illustrative. Use the team practice exercise before assigning reviewers. Suggest a correction.