Growth forconsumer brands,run by AI agents.

Alone: 4 of 6 steps

1. Set the rules
2. Check every row
3. Launch on Meta
4. Place creatives
5. Review the batch
6. Report back

Results

• 10 in under 5 minReview included.
• About 20 minPer launch with the agency.

Alone: 3 of 6 steps

1. Rate first outputs
2. Set brand rules
3. Make creatives
4. Check brand rules
5. Score them
6. Pick what to rework

Results

• About 100 a weekUp from about 10.
• About 30 minPer creative, down from about 3 days.

Founder’s estimate. No record.

Alone: 5 of 6 steps

1. Propose fixes
2. Check source data
3. Apply title fixes
4. Send the rest
5. Review
6. Re-test

Results

• 10 to 20%Lower CPA, not all from the agent.
• Checks addedAfter an invented attribute reached a live ad.

No check before each write yet.

Alone: 5 of 7 steps

1. Reconcile exports
2. Split the rise
3. Audit conversions
4. Reviewers check claims
5. Re-read live accounts
6. Build budget rules
7. Approve changes

Results

• 13 of 25Conversions set up wrong.
• 3.91xRise in cost per customer, partial month.

Alone: 7 of 9 steps, as designed

1. Propose the playbook
2. Find creators
3. Score them
4. Pick creators
5. Draft outreach (in build)
6. Approve, set the deal
7. Send (in build)
8. Set up tracking (in build)
9. Read results (in build)

Results

• ₹3,804 vs ₹9,401Per activation, creators vs ads.
• 24.7%Of first trades on 20.5% of paid spend.

Run by hand. No agents in these results.

Alone: 6 of 8 steps

1. Sort queries
2. Find gaps
3. Write pages
4. Ask LLMs monthly
5. Draft pitches
6. Approve
7. Send
8. Take the call

Results

• 163 to 483Daily clicks, December to May.
• 79 of 80Brand questions. Discovery: 1 of 80.

A product launch ran alongside.

Alone: 5 of 8 steps

1. Audit buyer searches
2. Pick channels
3. Build the list
4. Draft pitches
5. Approve
6. Send
7. Close
8. Audit placements

Results

• 23+Live pages. All paid, none earned.
• ₹4.25 lakhIn media. No retainer.

Alone: 5 of 6 steps

1. Collect notes
2. Find the ideas
3. Draft
4. Check the voice
5. Pick, edit, post
6. Log results

Results

• 735 to 19,311Impressions. Reach fell four posts running.
• 7 of 7Posts pass the voice check.

Alone: 7 of 9 steps

1. Cut the shoot
2. Master audio
3. Plan slides
4. Fetch images
5. Grade and caption
6. Render
7. Watch every render
8. Draft titles
9. Pick the packaging

Results

• 3Masters, longest 36 minutes.

One passed its checks with ten defects, so a founder watches every render.

Alone: 4 of 6 steps

1. Rank launches
2. Log who posted
3. Sample reach
4. Set the bar
5. Remove insiders
6. Sign off

Results

• 704Accounts added, 854 listed.
• 2,325Quote-posts across 149 launches.

Alone: 3 of 5 steps

1. Plan the calendar
2. Read it
3. Build campaigns
4. Check them
5. Approve sends

Results

• 5 minPer campaign, down from 45.

Alone: 6 of 7 steps

1. Set the rules
2. Count check-ins
3. Handle exemptions
4. Nudge
5. Compute scores
6. Write the message
7. Check every number

Results

• 1Wrong number caught live.
• 10+Weekly digests sent.

Never run for a brand.

Alone: 4 of 5 steps

1. Send check-ins
2. Answer replies
3. Pick what to fix
4. Write tests
5. Block bad restarts

Results

• 83 of 345Messages got a reply.
• 3 of 3Gate runs passed.

Personal trial, now paused.

Alone: 3 of 4 steps, as designed

1. Design flows
2. Take orders (in build)
3. Answer status (in build)
4. Win back (in build)

Results

• None yetVoice accuracy is the open problem.