RAZ AI · operating system
How I run three businesses on AI agents — the honest version.
Architecture, tools, approval gates, and the mistakes that cost me months. Filmed with my avatar; every word approved by me.
Chapters — 0:00 why · 2:10 the 3 parts · 4:30 lead-qualifier · 7:00 mistakes
The model
An AI agent has only three parts.
Something happens
A lead fills a form · an email arrives · a date approaches
One decision
Qualify · sort · draft · summarise · flag
Send, save or hand off
Anything client-facing waits for a human yes
In production
Agents currently running.
Lead-qualifier
Reads every enquiry, tags ready / needs-info / new, and drafts the reply for a human to send.
Read the system →Follow-up sequencer
Drafts the 2nd and 3rd-touch follow-ups for WhatsApp — where most conversions actually happen.
Read the system →Review collector
Asks, reminds and routes reviews; flags unhappy customers to a person first.
Read the system →Monday reporting
Leads, spend, source, ranges — every week, without anyone digging through dashboards.
Read the system →Content engine
Scripts → avatar render → schedule. One 60-minute approval → roughly 12 pieces a week.
Read the system →Ideas that failed the filter
About 80% of my automation ideas. Kept public for honesty — and to save you the months.
The graveyard →Two rules that make it work.
Agents prepare, humans approve.
Nothing a client or student sees is sent without a human yes. That one rule is why this works and most “AI automation” doesn't.
Revenue or 5+ hours a week — or it's parked.
The filter that killed 80% of my ideas and kept the 20% that run today.
The stack, kept honest.
Updated quarterly. Each is here because the system map asked for it — not because it was new.
- Claude / Cowork
- Meta Ads
- Google Ads
- WhatsApp Business API
- HeyGen avatar
- [CRM]
- [Automation layer]
- [Sheets / DB]
Want to learn to build agents like these?
It's in every Hashtag course and the 90-day program.