· 6 min read

Your AI Agent Still Needs You in the Loop

I watched an agent send the same follow-up twice. The customer replied to the first one. Nobody saw it.

By EZ4YouTech.com team

Agents copy and paste. They don't own the job. Put a person on the last click.

What broke on Tuesday

The agent didn't fail. The handoff did.

We pointed an agent at listing follow-up because the board was ugly. Red stickies. Names I didn't recognize. The agent drafted a decent first line. Then it posted without checking the first name against the file. I found out because the owner texted me from a job site, not because a dashboard pinged. There was no ping. There never is when you skip the last click.

The customer had already replied to the first note. Kind reply. Asking a real question about timing. That reply sat in a thread nobody watched. The agent sent a second note that sounded like we hadn't heard them. That's not a model failure. That's a workflow with no owner on send.

I still do this when I'm excited. I hook a tool. I go to lunch. I tell myself the hold state is 'coming next sprint.' Next sprint is a story we tell so we can leave the room.

What an agent is actually good at

Drafts. Recaps. Ugly first passes.

Voicemail to notes? Fine, if a person reads it. Estimate recap from a messy email chain? Fine, if someone who knows the job looks at the numbers. Photo to a punch list? Same rule. The second those notes go to a customer, a human has to hit send. I don't care how confident the model sounds. Confidence is cheap.

Where it dies: anything that commits money, dates, or names. Quotes. Change orders. Appointment texts that move a crew. I watched a shop almost dispatch to the wrong street because the agent grabbed an old address from a forwarded thread. The dispatcher caught it because he still reads. Barely.

Give the agent one queue. Not the whole office. If it can't show you the draft in a box you already live in, don't let it talk to customers. A Slack paste into a CRM field is not a workflow. It's extra steps with a nicer name.

The copy-paste loop nobody puts in the demo

Draft, paste, fix, paste.

I sat with a coordinator who had four tabs open. Agent writes. She copies. CRM rejects a character. She pastes again. Then she fixes the greeting because the agent said 'Hi team' to one person. That's not automation. That's a worse copier. The demo never shows this. The demo shows a green check.

Count the pastes for one afternoon. If the number doesn't drop after you add the agent, you bought a toy. I did this in a truck cab once because the office Wi-Fi was a joke. Not fancy. The count was honest.

Vendors will say 'agentic.' I hear 'it might send.' Until you have a hold, assume it might send. Turn auto-send off for a week and count how many drafts you'd have shipped raw. That number is your risk. If it's zero, fine. If it's twelve, you already knew.

A hold is not theater

It's the send button with a name next to it.

Named reviewer. Not 'the team.' A person who is actually in the building that day. If they're out, the queue waits. Waiting is cheaper than a second follow-up that sounds like you don't listen.

Log misses. Wrong name. Duplicate SMS. Quote with the old price. Don't log tokens. Tokens don't call you from a driveway. I keep a paper list when the tool's log is a mess. Paper still works.

  • One queue per agent.
  • Draft visible before send.
  • Human on send, named.
  • Misses written down the same day.

What I tried that wasted a month

Three agents. Zero holds. Lots of pride.

We split listing, intake, and reviews across three agents because it looked organized on a whiteboard. Whiteboards lie. Each agent needed a prompt, a login, a person who remembered the prompt. Nobody remembered. The listing one was the only one that ever ran. The other two became unread Slack channels.

Kill the extras. Keep the one queue that already hurts. Time five jobs by hand. Then let the agent draft. You click send. If you didn't save time, kill it the same week. Not after the quarterly review. After Friday.

Monday, if you're serious

Start with the pile on the desk.

Don't wait for a clean process. The pile is the process. Pick listing follow-up or estimate chase, whichever already makes someone stay late. Shadow that person for one hour. Count pastes. Put a hold on send before you buy another seat.

Coffee's cold. Phone's buzzing. That's the test. If the agent only works when the office is quiet, it doesn't work. I learned that on a Tuesday when three trucks called in at once. The agent kept drafting. Nobody clicked. The drafts expired. We did the follow-up at 7 p.m. like we always did. At least we didn't double-send.

Write the rule on a sticky: agent drafts, we send. Put it on the monitor that already has the other stickies. If it falls off, tape it. That's the whole product.

The inbox that ate the reply

The reply from the customer wasn't in the agent's world. It was in a Gmail label nobody checked because 'the agent handles follow-up now.' That's how a kind question about timing sat for two days. I found it while looking for something else. That's how most of our catches happen. Not monitoring. Accident.

We moved the label into the same box as the hold. Same screen. Same person. The accident rate dropped. Not to zero. To a number I can live with on a Tuesday. Zero is a story. Lower is a job.

What I tell a new dispatcher

Don't trust the green check. Trust the name on send. If you didn't type the name, look at it twice. Look at the street twice. Agents grab old streets from forwards. Forwards are poison. I treat a forward like a rumor until I open the file.

If the agent drafts in a voice you wouldn't use on a porch, rewrite the first sentence yourself. The rest can stay. First sentences are where we sound like a brochure. Brochures don't get jobs. Porches do.

Tools that didn't earn a second week

We tried a 'multi-agent standup' product. It made a summary of summaries. Nobody read it. The listing queue was still ugly. We cancelled on day nine. Day nine is allowed. Day ninety is how you pay for pride.

Another tool wanted calendar access on day one. No. Drafts only. Calendar is dates. Dates are money. Money is human. I keep saying that because I keep wanting to skip it when I'm behind. Behind is when you skip. Don't.

A week of holds in real numbers

I counted for five days. Not a study. A notebook. Forty-one drafts. Three I would have been ashamed to send. Two had the wrong first name. One had a time we'd already missed. That's five. Five is enough to keep the hold. I don't need a data team.

The ashamed ones weren't ungrammatical. They were empty. Empty is worse. Empty says we didn't read the last message. Customers punish empty harder than typos. Typos look human. Empty looks like a machine. Funny, given what we're using.

When the owner is the reviewer

Small shop. Owner is the named send. When they're on a roof, the queue waits. They hated waiting. They still hated the double-send more. Waiting won. The roof didn't fall off because a follow-up waited three hours. The double-send almost lost a job. That's the trade. I write it on the sticky so the argument dies faster next time.

If the owner refuses to be named, name the coordinator and give them the right to wait. Rights without a name are posters. Posters fade.

If You Only Do One Thing

Tape the four holds on the wall. Don't write a charter. Don't buy a seat. Don't schedule a town hall. Tape four lines. Then wait until a ticket tries to skip one. That's your first real meeting. It will be five minutes and it will be ugly. That's the meeting I want.

I keep a photo of that wall on my phone. Not because I'm proud. Because I forget. Forgetting is the product. The photo is a product too. A sad one. It works.

Next step

Start with one queue, not a stack of agents.

Browse More AI Insights All articles