Discover the break. Propose one fix. A human approves any money write. Slack is the operator surface. That is the billing agent I run today. It is not a chatbot on a spreadsheet.
I'm Michael Thomas Zajac. Hermosa Beach. This is the same operator work I still do in RevOps: sit with the people who own the book, find where the process leaks, build the system they will actually use.
What was broken?
Billing breaks are ugly because the cost of a wrong write is real money. Teams either keep a human in every cell, which does not scale, or they let a model post a fix, which is how you invent a second set of books.
Automation without a human gate is how you burn a domain, a brand, or a book. On billing, you burn cash.
What did I build?
A control proof. The app is MCP-layered. Embedded cloud agents do three jobs. None of them get the checkbook.
Triage. Agents classify the break. What kind of miss is this. Who owns it. What evidence do we already have. No write.
Business logic remediation. Known rules propose one fix. One. Not a menu of maybes. The operator sees the break, the rule that fired, and the proposed write in Slack.
RAG-architected learning. The third layer. Counts, rules, and alerts that should be added to business logic get pulled from what the agents already saw. That learning feeds back into layer two so the next pass uses a better rule, not a longer prompt.
MCP is how the agents get tools and context. The human is how money moves.
What is the number?
There is no recovery dollar on this one. The control is the number. One proposed write. One human approval. That is the gate.
What should you do Monday?
Pick one workflow that can write money. Split it into triage, propose, approve. If your agent can post the journal entry without a human, turn that off. Then add one count, one rule, or one alert from last week's misses back into the logic. That is the third layer. That is how the system gets smarter without getting a pen.