Every agent I build follows one rule: it drafts, and a person decides. Nothing is sent and no CRM record changes until someone approves it.
That sounds slow. It isn’t, because the agent does most of the work before a person looks at it. My rule of thumb is 80/20. The agent does about 80 percent. I do the other 20 percent, which is checking it for made-up facts and weak writing before it leaves.
How a run flows
- A trigger starts the run: a schedule, a new deal, a call transcript or a question.
- The agent reads one shared configuration and pulls the records it needs.
- It drafts the result with the evidence behind each claim.
- The draft waits in an approval queue.
- A person approves, edits or rejects it. Only then does anything reach the CRM, an inbox or a customer.
Three details that make it work
One setup, shared. A single interview captures the target customers, buyer personas, competitors, field mapping, email voice and guardrails. Every agent reads it, so no agent asks the same question twice.
Evidence on every claim. Each statement an agent makes carries a quote, a link or a record it came from. If I can’t see where a claim came from, I don’t send it.
Changes shown as a table. Proposed CRM edits appear as a table of before and after. I apply them after I read them, not before.
The catch
The approval step is work, and it is the work that matters. An agent will sometimes state something that sounds right and isn’t. A person who reads the evidence catches it. A system that skips that step ships it.
You can see the full list of agents I’ve built and how each one is controlled on the AI agents page.
I’m Seth Levanen. I find the AI shortcuts that keep the quality and drop the busywork. Email me if you want to talk about it.