Your AI Agents Need a Code of Conduct. Here's What Goes In It.

When you hire a new project manager, you don't just hand them the keys to the truck and the login to your job software. You tell them how you run things. Don't promise a homeowner a finish date you can't hit. Don't approve a change order without a signature. Call me before you spend more than a certain amount. Treat the customer's house like it's your mother's.

Most of that never gets written down. It lives in how you train people, in the stories you tell, and in the times you've pulled someone aside after a job went sideways.

AI agents don't pick up any of that by osmosis. If you want an agent to work the way you work, you have to write it down. That's what a Code of Conduct is for.

Why an agent needs one

An agent that only answers questions doesn't need much in the way of rules. An agent that acts is different. Once it can log in to your job software, draft estimates, send follow-ups to adjusters, or track expenses, it's making small decisions all day long. Every one of those decisions either matches the way you'd do it or it doesn't.

Without written rules, the agent fills the gaps with its own best guess. Sometimes that guess is fine. Sometimes it sends a homeowner an email you'd never have sent, or reports a job as done when it isn't.

A Code of Conduct closes those gaps before they cost you a customer.

There's a second reason. If you run more than one agent, as I do, the Code of Conduct keeps them all on the same page. My main agent, my editor, and my financial agent all work under the same charter. When one hands work to another, they're playing by the same rules.

What it should cover

Mine started as a one-page charter and has grown as I've learned. Here's what I'd put in yours.

Loyalty to you. The agent works for you and your business, period. When two goals conflict, it resolves them in your favor and tells you about the conflict.

Care. Do the work carefully. Check it before calling it done. A finished job means tested and working, not "should work."

Confidentiality. Customer names, claim details, insurance information, and your numbers stay private. Keep a written list of every outside service that can see your business data, so you always know where it lives.

Nothing goes out without your OK. This is the big one. No emails, texts, or posts to customers, adjusters, or subs unless you've reviewed them. No spending money. No deleting anything. No sharing files. The agent drafts and you send.

Candor. Tell the truth, especially when it's bad news. If something broke, say so. If it doesn't know, say it doesn't know. I'd rather hear "I couldn't verify that" than get a confident wrong answer.

Every claim gets a source. If the agent tells you a supplier price went up 8 percent or a permit takes three weeks, it should be able to show you where that came from. If it can't, it should say so.

Separate what it did from what you did. When it reports results, it should be clear about which work was the agent's and which was yours. That matters when you're judging whether the agent is actually earning its keep.

Measure before and after. Before an agent takes over a task, write down how long it takes now and how often it goes wrong. Give the agent credit only for change you can actually measure.

Plain words. No hype, no jargon, no fluff. If it can't explain something in shop English, it doesn't understand it well enough.

Recovery. Know where your work is saved and how to get it back. I learned this one the hard way when an early agent built me an app and never saved the data. Two days later it was gone.

How to put one in place

Write it like you'd talk to a new hire. Skip the legal language. Short, plain statements work best, for the agent and for you.

Put it where every agent reads it. Most agent platforms let you give each agent standing instructions or a shared file. Put your Code of Conduct there so it applies every time, not just when you remember to mention it.

Start small and add as you learn. You don't need fifty rules on day one. Start with the big five: loyalty, care, confidentiality, no sending without approval, and candor. Every time an agent does something you didn't like, ask yourself whether a rule would have prevented it. If so, add it.

Review it like any other policy. Once a quarter, read it again. Cut what doesn't matter and tighten what does.

Hold the line. If an agent breaks a rule, fix the instructions right away. Don't let it slide just because the result turned out fine that time.

The bottom line

You wouldn't put a new hire on a water loss with no training and no rules. Don't do it with an agent either. A one-page Code of Conduct is the cheapest insurance you'll ever buy for an AI workforce, and it's the difference between an agent that guesses and an agent that works the way you do.

John H. Hodsdon writes about practical AI for home services businesses. Find free guides and training at agenticsolutions.biz.