For BMG, prepared by Peter Utekal
A working model: small human-led pods, each supported by Grok Bot assistants and Cursor cloud agents, with people approving anything that touches money, brand or the client.
Scroll to move through the sections.
2 of 9
At 30+ clients, most of the team's week goes into the same repeat work for every account. That work can be handled by assistants, so people spend their time on judgment and client relationships.
Every ad account needs spend, pacing and CPA checked each morning. With 30+ accounts this is hours of looking before any decision is made.
Meta's delivery rewards a steady supply of distinct creative (Meta on creative diversification). Briefs, variants and QA have to keep up for every client.
Weekly reports and monthly reviews follow the same structure each time. Learnings from one client rarely reach the others.
The idea: assistants do the pulling, drafting, checking and tagging. People keep four decisions: what we say, what goes live, what we spend, and what we tell the client.
3 of 9
Three pods of about 10 clients each, coordinated by an agency COO bot that reports to BMG leadership. Shared bots and Cursor cloud agents serve every pod.
Swipe sideways to see the full diagram.
Each pod is a Pod Lead (human) and seven bots that cover the full client cycle. The pod is the unit we copy when the agency grows.
Work that is the same for every client sits outside the pods: selling and onboarding, finance, and team operations.
Build and maintain the code: scripts against the Meta API, the dashboard across all accounts, and the client portal.
4 of 9
Work moves left to right from brief to live ads, then the Analyst feeds results back to the Strategist. Black circles are human gates.
Swipe sideways to see the full diagram.
Pull data, write briefs and scripts, produce variants, check policy and brand, prepare launches and budget changes, tag results, draft reports.
Approves at the four gates, talks to clients, and decides when something looks wrong. One Lead covers about 10 clients.
Each client has a memory file: offer, audience, brand rules, past winners, feedback. Every bot reads it before working on that client.
5 of 9
Five recurring routines. Each one is a scheduled bot routine with a clear hand-off to the Pod Lead.
Each morning the Analyst pulls spend and CPA for all 10 accounts. The Media Buyer proposes budget changes and pauses. The Lead approves anything over the limits (G3).
Fatigued ads are flagged. The Strategist writes new angles (G1), Production makes variants, QA checks them (G2), and the Media Buyer launches a test.
Winners are tagged by hook, format and audience and added to the Strategist's library. The library is shared across all pods, so a win for one client informs the others.
Bots draft the weekly report and the monthly review. The Lead sends them (G4). Client feedback is logged into that client's memory.
First, ad account, page and business access is granted and checked, and the pixel, Conversions API and key events are verified. Next, the Pod bot writes the client's brand memory from their site, past ads and the kickoff call. The first test batch is live by day 7. The rest of the two weeks goes into reading first results with the client and confirming targets.
6 of 9
Suggested starting targets, tune to the agency's benchmarks.
| Role | What we measure |
|---|---|
| Pod bot | Client requests answered the same day. 100% of briefs complete before production. |
| Creative Strategist | 3 to 5 new angles per client per week. Test win rate of 20% or more. |
| Creative Production | 10+ variants per client per week. Turnaround of 48 hours or less. |
| Policy & Brand QA | Under 5% of ads rejected by Meta. Zero account warnings. |
| Media Buyer | Spend within 10% of plan. CPA or ROAS at or better than the client target. |
| Analyst | Fatigue flagged within 24 hours. 80% or more of alerts lead to an action. |
| Reporting | 100% of weekly reports sent by Monday. |
| Pod overall | 8 of 10 clients on target. Churn under 5% a month. Revenue per head. |
7 of 9
Bots prepare, people decide. These four points always need a person.
Any new angle or offer is approved before production starts.
Anything that goes live with the client's brand. Client sign-off as well, for clients who want it.
New campaigns, budget changes over about 20% or a fixed amount, and turning off a top spender. Above a higher limit, it goes to BMG leadership.
Every client message and report. Bots only draft, the Lead sends.
An account warning or ban. A client off target for 2 or more weeks. Any sign of churn risk.
8 of 9
Existing tools we would plug in, plus the skills we would write for the agency. Everything listed below exists today.
| Stage | Best picks | We write |
|---|---|---|
| Pod bot |
|
Client memory format and brief template |
| Creative Strategist |
|
Hook and angle framework, fed by our winners library |
| Creative Production |
|
Variant spec: sizes, safe zones, naming |
| Policy & Brand QA |
|
Agency QA checklist plus each client's brand rules |
| Media Buyer |
|
Budget rules that match gate G3 |
| Analyst |
|
Tagging taxonomy: hook, format, audience |
| Reporting |
|
Weekly report and monthly review templates |
| Onboarding |
|
14-day onboarding checklist |
Alternatives worth a trial: Adspirer, Madgicx MCP, Ryze AI (account management), Atria MCP (ad library), Higgsfield MCP (video). Third-party skills should be read and tested on one account before use.
9 of 9
Start with one pod. Copy it only once it works.
One Pod Lead and about 10 clients. Connect ad accounts read-only first, write each client's memory file.
Daily pacing, weekly creative, weekly learning and the client loop. The Lead approves everything at first, then loosens limits where the bots prove reliable.
Compare the pilot pod to the rest of the agency on the targets in section 6. Fix what does not work.
Roll the same bots, skills, memory format and gates out to the other pods. The shared learning library connects them.
Prepared by Peter Utekal.