Running an AI Team Like a Department
I designed and run a team of AI models (Claude, GPT, Gemini and open-source models) that research, write, review and produce video every day, on a self-managed server for about USD 85 a month, cutting the cost of an explainer video from about VND 650,000 to a few thousand dong.
- Milestone: The whole AI team runs for ~USD 85 a month
- Decreased: Cost per explainer video fell from ~VND 650,000 to a few thousand dong, voice, script and editing included
- Milestone: 4 AI team members, each using the right model for the task
Why I built it
I do not have a coding background. What I do have is a long dislike of repetitive work, and a clear view of how good marketing gets made: research, writing, critique, production, review.
In 2026, I asked a simple question: could AI do that work the way a well-run department does, with roles, handovers and a manager, instead of one chatbot doing everything?
How the team works
Four members, each with a role. Claude, GPT, Gemini and an open-source runner each take the work they do best. Every task gets the right model, not the most powerful one: lighter models for routine work, the strongest models only where quality matters.
One shared memory. The whole team reads from the same knowledge base, so context is never lost between tasks or members.
A manager, and a human with the final say. An orchestrator plans the work, calls the members and writes the result back. I approve key decisions from my phone, in a simple web app where every project looks like a group chat.
Built-in critique. For a script, all four members write their own version. I choose one, the other three critique it, and the author finalizes it. That beats any single model working alone.
What it produces
- Topic research, collected continuously into a topic bank for my channels.
- Scripts and posts, written, critiqued and finalized by the team.
- Finished videos in one click: once I approve a script, the team creates the voice-over and the edited video by itself.
- Websites and tools, built with AI from design files.
Results
| Area | Result |
|---|---|
| Running cost | about USD 85 per month for the whole team, on a self-managed server |
| Cost per explainer video | from ~VND 650,000 to a few thousand dong, previously a writer, a voice talent and manual editing |
| My role | deciding what to make and judging quality, not doing the repetitive steps |
What I learned
Not everything should be automated. When I let the members coordinate freely on a coding project, the process became too complex, so I went back to a simpler, more manual flow for that work. Knowing where to stop automating is part of the design.
AI does not replace judgment. It multiplies the value of good judgment, and the cost of bad judgment.
The same principle drives my work with businesses: I apply AI where it changes results, such as the AI-powered lead system behind 10+ weekly site visits for a luxury real-estate developer.
