Where should an AI workflow stop?

My rule is simple: AI can draft internally, but anything client-facing gets my review. That includes emails, page content, email newsletters, and entire websites.

I don’t need to watch every keystroke while a draft comes together. The boundary that matters is the moment the work would leave my workspace and represent me or my client. A polished draft can still contain a made-up fact, a promise nobody approved, or a link that goes nowhere. If it goes out under my name, I need to be able to stand behind it.

What my review caught on a recent rebuild

I used an AI coding agent on a website rebuild for a recovery community organization. The examples here come from the preview, while final changes and client sign-off were still pending. Three moments from my review show what the rule looks like in practice.

The homepage photo. The client sent new photos of their activities, and I asked the agent to swap them in where they fit. It took that further than I meant and replaced the homepage hero photo too. I told it, “Put the original hero headshot back. I don’t want that one to change.” It restored the original photo and the matching share image (the picture that shows up when someone posts a link to the site). Then it wrote that choice into the project’s design notes so later updates leave it alone. Permission to refresh supporting images isn’t permission to change the image that defines the page.

The mobile menu. The agent built a dropdown menu for phones. It worked, but I wanted the drawer menu I already use on 21adsmedia.com, so I asked for that instead.

The recovery phase cards. The site walks visitors through the program in phases. The agent’s first version showed those phases mostly as text. The old WordPress site had artwork for each one, so I pointed the agent back to it, asked for the original icons and images, and asked for that card to be reused across the site as part of the branding. Those revised cards made it into the preview.

Each of those changes made sense on its own. What the agent couldn’t know without me was which parts of the old site were the client’s identity and which were fair game. My goal was to modernize how the site is built without casually replacing the look the client already established or the design references we’d already approved. That call is the part I don’t hand off.

The same rule at different scales

Work The agent can prepare I review before it is client-facing
Email A reply in the right thread Recipient, facts, commitments, tone, and send
Newsletter Subject, copy, links, and layout Claims, audience, rendering, links, and send
Page content Draft copy and structure Accuracy, client voice, calls to action, and publish
Website A working build or preview The whole site, with the same scrutiny I would give a manual build before launch or handoff

The review changes with the stakes. A short scheduling reply may take a quick check of the calendar and the recipient. For a website, I review everything, just as I would if I built it by hand. I use the actual site, check it on different devices, read its content, and test how its pages and interactions work. The client collaborates on content, usually in a Google Doc, and signs off before launch. “Human review” means looking at what the visitor will see and do, not clicking an approve button.

This rule doesn’t make AI passive. It can gather context, compare options, write, revise, and assemble a preview. The draft can get close. The final client-facing decision stays with the person accountable for it.

What that means if you work with 21 ads media

I use AI in the production process because it helps me explore, build, and revise. I still take responsibility for the result. If 21 ads media builds your site, expect me to review it as thoroughly as I would a manual build, check the experience across devices, and bring you into the content review before launch. You sign off on the final site. That includes protecting the parts of your brand people already recognize. The point is room for better problem solving and careful iteration, not trading quality for more output.

If you have a site or workflow that needs to work for the people using it, tell me what it needs to do. We can start there and decide what’s worth building.

A boundary you can copy

For each workflow, define two things in plain language: what the agent may prepare, and what action would make that work visible to someone else. Have it stop before that action, show you the draft or preview, and name any facts it couldn’t verify. If there are things it shouldn’t change (a logo, a hero photo, a design you already approved), write them down where the agent will see them every time. Then review the result in the form the recipient will experience.

This week: Pick one task you hand to an agent and write down exactly where it should stop. For a website update, that could be a working preview for you to test on your phone before anything goes live.