Skip to article

How Do You Keep an AI Agent From Sounding Like AI?

A draft can pass every word check and still sound like a machine. Two of my corrections, and how to check an AI agent's drafts for it.

By Jacqueline V. TwillieNovember 6, 20262 min readGuide
Quick answer

Teach your AI agent from emails you've sent, put the shapes to avoid in its voice rules beside a fixed example, and have a second AI agent check each draft before you read it. A word check alone clears drafts that still sound like a machine.

In July a draft passed every word check I had, and I sent it back: "remove all the AI tropes."

One line explained itself.

After that, my voice rules got a section on how a sentence is built, and it opens with a pair. The first line reads, "Create your account and have it open in a browser tab before we start, so the first ten minutes go to building instead of sign-up screens." The fix reads, "That's where we build. Have it open before we start." The second line is in my reminders.

The old line gives an instruction and then explains why it matters, and it made up the ten minutes. A word list can't flag either one. My AI agents read the pair and the rule it breaks. Yours won't know the shape until you show it a pair: the line as drafted and the line as it should read.

The emails passed too.

On November 5 I read eight test sends from a campaign. A word check and a reading pass had cleared them about twenty minutes earlier. I said, "the emails have far too many AI tropes."

The campaign shared what people had told me about my build sessions. It kept putting a number next to what they said. The numbers came from real answers, and in a letter to friends they'd still read like a report. A note about how many people had answered came with them. The same ask showed up in five pieces.

My emails open with a hello or the thing you're about to do, never on a statistic. When I ask for something, I leave a way out: "Can't join live? Reply here and I'll make sure you're set up for what's next."

The numbers in my emails are mostly a start time or a discount code. Give your AI agent a few emails you've sent and have it build the next one from them with the details swapped. Then add this to its voice rules: in a letter, keep the numbers to the ones that change what the reader does, and cut any sentence that explains how you counted.

Who reads the draft before you do?

I call mine the Brand Voice and Fact-Check AI Agent. It checks every draft against my voice rules and my sources before it reaches my desk. You still read the draft. Start with the last one your AI agent wrote and count its numbers.

More on Ownership for Mid-Career Professionals