From meeting transcript to action items, with a human check on every line

How the workflow runs

The same four steps for every meeting, so the output stays consistent no matter who runs it.

1. Paste the transcriptDrop any meeting transcript, from a status meeting to a design review, under a saved prompt.
2. AI drafts the documentsUnclear owners become UNASSIGNED and missing dates become TBD. Nothing is guessed.
3. A person reviews every lineOwners, dates, and decisions are checked against source quotes with an 8-point checklist.
4. Approve and sendReviewed notes go to the team and leadership, and action items flow into Jira.

Prompt library

Each prompt tells the AI to use only what was said, quote its sources, and mark gaps instead of filling them.

How I test accuracy

Instead of trusting the output, I built an answer key for all three meetings and score every AI run against it. Each meeting hides traps that a careless summary falls into.

Baseline prompt versus guarded prompt

Summarize this meeting and list the action items with owners and due dates.

Use only what is stated. If no owner was clearly named, write UNASSIGNED. If no due date was stated, write TBD. Never guess. Quote the source line for every item.

Team adoption plan

A one-page operating procedure so a program team can use the workflow safely, not just one person.

Rollout

1. Pilot on one recurring meeting for two weeks, reviewing every draft against the transcript.
2. Track the correction rate and time saved each week in the correction log.
3. Expand to change control and design reviews once owner and date accuracy hold steady.
4. Review the prompts monthly and update them based on the most common corrections.

Automated follow-ups

Reviewed action items go into a Google Sheet. A small Apps Script runs every morning, reminds owners of items due within two days, and emails the project manager a list of anything still UNASSIGNED or TBD.

function sendFollowUps() {
  const sheet = SpreadsheetApp.getActive().getSheetByName('Action Items');
  const rows = sheet.getDataRange().getValues().slice(1); // skip header row
  const pmEmail = 'your.email@example.com';
  const today = new Date();
  const gaps = [];

  rows.forEach(([action, owner, due, status, ownerEmail]) => {
    if (status === 'Closed') return;
    if (owner === 'UNASSIGNED' || due === 'TBD') {
      gaps.push('- ' + action + ' (owner: ' + owner + ', due: ' + due + ')');
      return;
    }
    const daysLeft = (new Date(due) - today) / 86400000;
    if (ownerEmail && daysLeft <= 2) {
      MailApp.sendEmail(ownerEmail, 'Reminder: ' + action,
        'Hi ' + owner + ', this action is due ' + new Date(due).toDateString() + '.');
    }
  });

  if (gaps.length) {
    MailApp.sendEmail(pmEmail, 'Action items needing an owner or date', gaps.join('\n'));
  }
}

Sheet columns: Action, Owner, Due, Status, Owner email. Set a daily time-driven trigger in Apps Script to run it each morning.

Automatic Jira tickets

Once action items pass human review, a second script turns each one into a Jira task. Owners are assigned automatically when they have a Jira account, and anything still UNASSIGNED gets a needs-owner label so it can't be lost.

1. Review in the sheetReviewed action items land in the Action Items sheet with owner, due date, and status.
2. Script creates ticketsEach open row becomes a Jira task with summary, due date, assignee, and labels.
3. Ticket key written backThe new key, like MTG-12, is saved in the sheet, so the same item is never ticketed twice.
4. Gaps stay visibleUnowned tasks carry a needs-owner label, ready for a saved Jira filter the PM checks daily.

Field mapping

The script

const JIRA_BASE = 'https://your-site.atlassian.net';
const PROJECT_KEY = 'MTG';

function createJiraTickets() {
  const props = PropertiesService.getScriptProperties();
  const auth = 'Basic ' + Utilities.base64Encode(
    props.getProperty('JIRA_EMAIL') + ':' + props.getProperty('JIRA_TOKEN'));
  const sheet = SpreadsheetApp.getActive().getSheetByName('Action Items');
  const data = sheet.getDataRange().getValues();

  for (let i = 1; i < data.length; i++) {
    const [action, owner, due, status, ownerEmail, ticketKey] = data[i];
    if (status === 'Closed' || ticketKey) continue; // skip closed rows and rows already ticketed

    const fields = {
      project: { key: PROJECT_KEY },
      issuetype: { name: 'Task' },
      summary: action,
      labels: owner === 'UNASSIGNED' ? ['meeting-action', 'needs-owner'] : ['meeting-action'],
      description: { type: 'doc', version: 1, content: [{ type: 'paragraph', content: [
        { type: 'text', text: 'From reviewed meeting action items. Owner: ' + owner + '. Due: ' + due + '.' }] }] }
    };
    if (due instanceof Date) fields.duedate = Utilities.formatDate(due, 'UTC', 'yyyy-MM-dd');
    const accountId = ownerEmail ? findAccountId(ownerEmail, auth) : null;
    if (accountId) fields.assignee = { id: accountId };

    const res = UrlFetchApp.fetch(JIRA_BASE + '/rest/api/3/issue', {
      method: 'post', contentType: 'application/json',
      headers: { Authorization: auth }, payload: JSON.stringify({ fields }),
      muteHttpExceptions: true
    });
    if (res.getResponseCode() === 201) {
      sheet.getRange(i + 1, 6).setValue(JSON.parse(res.getContentText()).key); // write ticket key back
    } else {
      Logger.log('Row ' + (i + 1) + ' failed: ' + res.getContentText());
    }
  }
}

function findAccountId(email, auth) {
  const res = UrlFetchApp.fetch(
    JIRA_BASE + '/rest/api/3/user/search?query=' + encodeURIComponent(email),
    { headers: { Authorization: auth }, muteHttpExceptions: true });
  const users = JSON.parse(res.getContentText() || '[]');
  return users.length ? users[0].accountId : null;
}

Works with your team's tools

The workflow does not depend on Jira. Transcript, AI draft, human review, and the tracker sheet stay the same. Only the last step changes: which tool receives the reviewed action items. Jira is built and tested. The others follow the same pattern with a different web address and field names.

No-code option: Zapier, Make, or Power Automate can send new tracker rows to any of these tools, so teams can adopt the workflow without maintaining scripts.

See it working

I ran the full workflow end to end on a free Jira Cloud site, using the 15 reviewed action items from the answer key. These screenshots are from that run. Personal email addresses are blurred.

15

Jira tickets created by the script, KAN-4 to KAN-18

5

gaps flagged to the PM: 4 unowned tasks and 1 missing date

1 click

to run everything from a custom Meeting Actions menu in the sheet

Human review rules

AI drafts the documents. A person approves them. These rules apply before anything is shared.

• Verify every ownerRaising an issue is not the same as owning it.
• Verify every dateDates must come from the transcript, or they become TBD.
• Check decisions against quotesOptions that were only discussed stay out of the decision log.
• Chase every gapEach UNASSIGNED or TBD item gets a follow-up within one business day.
• No invented factsAnything the transcript doesn't support is deleted, even if it sounds right.
• Protect sensitive informationNo patient data, proprietary designs, or confidential company material goes into an AI tool. Approved tools only.
• Keep a correction logEvery edit to an AI draft is recorded to improve the prompts.
• AI notes are not quality recordsDesign decisions and change approvals are recorded through formal document and change control, not AI drafts.

The three test meetings

A fictional surgical robotics software release (Aurora SR-2, release 2.4), written with vague owners and shifting dates on purpose, so the review rules have real errors to catch.

Why it matters