How I use AI

“If a reviewer asks, ‘What did you mean by this line?’, it’s not acceptable to reply with ‘Oh sorry, AI wrote that, just ignore it.’”

It’s what Sophie Alpert wrote in Clay’s AI writing policy. When I read it, I was like, “Wait, WTF, people actually do that?” It’s crazy to me that anyone would leave AI copy unchanged and then use the model as an excuse.

As the copywriter, my job is to bring the expertise clients need to turn their internal thinking into something customers can understand and act on. I have to stand behind every idea, every emphasis, and every sentence. AI can never take responsibility for any of it.

Deciding what to emphasize, how to structure an idea, and what to leave out is all part of understanding and clarifying what you mean.

In messaging work, those decisions should come after research into the people we’re trying to reach: what they’re trying to achieve, what they’re struggling with, what they want instead, and what might stop them from acting. Then comes positioning and messaging strategy. The copy is only downstream of all of that judgment.

Looking exclusively at the final product can be misleading. A reader might skim a page in five seconds when the work behind it took two or three months. They don’t see the interviews, competing directions, weaker variants that never made the cut, or rounds of iteration. They see the small amount of information that survived all this pruning. In one of my fave books of all time, The User Illusion, Tor Nørretranders calls the information we deliberately discard while shaping a message exformation. Concise copy is just the residue of a lot of research, selection, and subtraction.

We never write copy to show off (clear > clever, remember?). Our goal is instead for the right reader to understand it, realize that it speaks to their exact situation, and take the next step. An AI model can predict and put together language patterns, but ask it to own its decisions or even explain why it decided to do X over Y, and you’ll always get new BS excuses on top of BS excuses. You, as the human directing the work, have to make that call.

Reading Clay’s AI writing policy reminded me of a Paul Graham essay I read a while ago about what good writing actually is. His argument is that getting the writing right often means developing the ideas properly: drawing the conclusions that matter and exploring them to the right level of detail. For example, when a passage feels awkward, rewriting it will likely show you what you didn’t think about deeply enough or clearly enough. As Graham puts it, the writer is the first reader. If we jump too quickly from prompt to final copy, we are basically skipping the entire part of the work that helps us empathize with and understand our ICP.

There is some research that points in the same direction, with important limits. In a CHI 2025 study of 319 knowledge workers, Hao-Ping Lee and colleagues collected 936 first-hand examples of GenAI use at work. People with higher confidence in GenAI reported less critical thinking, while those who trusted their own ability to perform a task reported more. Participants also said their critical-thinking effort shifted toward verifying information, integrating responses, and overseeing the task. The study was self-reported and associational, so it doesn’t prove AI causes people to think less. But it does suggest AI changes where, across the different stages of our work, we now think independently—if we do at all—which is the biggest risk.

The fair counter-reading is that AI can genuinely help people think and write. So, inspired by the team at Clay, I wanted to sit down and think through how I really use AI at Conversion Alchemy, so I could share it openly with clients and peers. You can check out my new How we use AI page here.

I use agents throughout all my work. They can provide options I hadn’t considered, challenge an assumption, organize messy material, catch errors, handle repetitive information work, and help me get from a blank page to something I can use to think and write even deeper.

For example, something I’ve realized while writing that page is that speed and scale require more accountability and judgment, not less. Especially because I’m a solo operator working with a team of AI agents. If more of everything I work on passes through me, then I need to be an even better evaluator and manager.

Which means I still have to verify the data, decide which ideas are worth exploring, integrate them into a coherent argument, remove what doesn’t help, and approve the final result. Producing more words isn’t automatically more useful either. Length and detail should follow your audience, channel, level of awareness, sophistication, and decision-making process, not the model’s ability to keep vomiting out words.

The standard I’m using is simple: can I explain why each important idea earns its keep, what supports it, what I deliberately left out, and what decision we’re helping our reader make? If I can’t, my agents and I have got more work to do.

Your task for the weekend: sit down and write how you actually use AI in your work. You don’t even have to publish it. The act of writing it down will give you a ton of clarity that you can then carry into what you do.

But hey, if you already have an AI policy page, I’d love to check it out.

Discovery

Let your work decide what you delegate to agents

“Is the model smart enough?” isn’t the most useful question for deciding what to delegate to an AI agent. PostHog’s agent-autonomy framework uses two more practical tests: is the work easy to check, and is a mistake cheap to undo?

When you can both verify and reverse your work easily, you get into self-driving mode, which is where agents are truly magic. But, on the other hand, when your work is ambiguous or expensive to fix, keep your helpers as assistants. That maps well to writing: proofreading can run freely; choosing your market position cannot.

Ask for feedback early

Jason Freedman’s Thirty Percent Feedback recommends telling whoever has to review your work whether you are 30% or 90% done. At 30%, feedback can challenge your direction. At 90%, it naturally narrows to the details.

That’s useful for agents too. Have them bring up the POV, argument, assumptions, and proposed structure while their work can still change materially. If you only enter the loop when a polished but mostly slop draft is “done,” it’s already too late.

Resonance