“I rejected 23 of 26.” Why human in the loop AI beats letting an agent hit send.

Why do AI agents need a human approval step?

AI agents need a human approval step because a draft can be clean, polite, and still wrong. An agent doesn’t know what you promised in the meeting, what you’re attaching, or which client hates a long email. It fills the gap with a good guess.

For a benefits agency, a good guess sent to a client is a problem you’ll spend a week undoing: a wrong renewal date, a price nobody quoted, a promise nobody made.

What do most agencies try first?

Most agencies try one of two things. Some let the AI send on its own and hope the first few emails set the pattern. Others ban AI from email entirely and keep typing every reply.

Both cost you. Auto send puts guesses in front of clients. A ban throws away the part that works: the reading, the sorting, and a first draft waiting when you sit down.

Here’s what my 23 rejections were about. After a meeting, I send a reply with the deck we went over attached, and I hadn’t built the drafter to attach a document. So I wrote my own reply, then opened the AI draft to check if I missed something. If I had, I added it to my draft. The AI draft became my checklist, and the only thing left to do with it was reject it.

Wanna fix it?

AI that could send the wrong thing gets fixed with human approval and guardrails: agents that can read and write drafts but never send, written rules for what a draft can never say, and a person who approves every draft that goes out.

1. Take away the send permission. My reply agents can read the inbox and write drafts. A person sends. It takes a long time to train a reply agent, and until it’s trained, drafts are where it lives.
2. Write voice rules. The guardrail that has stopped the most mistakes for me is a list of what can never be in a draft: banned phrases, no dashes, a set length, my sign off.
3. Ban invented facts. My reply drafter can never make up a time, a price, a case study, a client name, or a result. If it doesn’t know, it leaves it out.
4. Make approval one click. Approve, edit, or redraft with a note like “shorter.” A slow queue gets skipped.
5. Earn automation one issue at a time. A few common issues with a clear template can run with less review. Everything else, a person’s judgment is the bar.

What does it look like once it’s running?

Once human approval is running, the inbox is read and sorted before you open it. Real replies sit in a queue with a draft attached. You approve, fix, or redraft. Nothing goes out on its own.

The same idea runs underneath everything else. My mailboxes are opened read only, so an agent can’t even mark a message read. My memory system proposes new facts and waits for approval before saving them. The weekly cleanup can change 15 memory notes at most, never deletes anything, and has one standing rule: when unsure, don’t act.

Can you do this yourself?

Yes. If your AI email tool has a draft only setting, turn it on. Then write a one page list of what a draft can never contain, and paste it into the tool’s instructions. Review every draft for a month and keep a tally of what you changed.

That tally tells you when a reply type is ready for less review. It stops working when drafts come from five places and nobody knows which rules apply to which one. One inbox, turn on draft only. Agents writing across your CRM, email, and files, build the approval queue.

Quick answers

What is human in the loop AI?

An AI setup where agents draft and a person approves anything that goes to a client.

What are AI guardrails?

Written rules and permissions that limit what an agent can say or do, like no sending and no invented prices.

What is an AI approval workflow?

A queue where every draft waits for a person to approve, edit, or redraft it.

Should AI send client emails on its own?

Only for a few common issues with a clear template, after the agent has been trained on them.

What guardrail stops the most mistakes?

For Chris Cordon, voice rules: a written list of what a draft can never contain.

Has DoBetter built this?

Yes. Chris Cordon’s CRM drafted 26 meeting follow ups; he sent 3 after editing and rejected 23.