What Can AI Actually Do in WordPress? What I'm Hoping For, What I Expect

I build WordPress sites for a living, and I still haven’t let AI touch a single client’s site.

The last few months have been heads-down work on real websites for real businesses while also testing, learning, and perfecting strategies for using AI in other parts of Alt Creative’s business operations. The more I experiment with AI, the more interested I am in using it to bring value to my clients’ websites.

So before I sit down and put it through its paces, I want to write down two things: what I’m hoping AI in WordPress can do, and what I expect it’ll take to get there. I don’t think those are close together. The hope is a clean, finished outcome. The reality, near as I can tell, is a lot of trial and error first.

What can AI do inside WordPress right now?

As of 2026, AI in WordPress works on three levels: it can generate content like drafts and alt text, build pages and layouts from a prompt, and, with WordPress 7, connect to your site as an agent that makes changes for you after you approve them. The third level is the genuinely new part.

AI inside WordPress now spans three levels:

• it can write content for you,
• it can build pages and layouts from a prompt,
• and, as of this year, it can connect to your site as an agent and make changes on its own with your approval.

WordPress 7 shipped the foundation for that last part, which is the new piece that matters.

The first level is old news. For a while now you’ve been able to generate blog drafts, meta descriptions, and alt text without leaving the editor. Alt Creative has already been using it in this capacity, but it’s never automated. We use the human-AI-sandwich approach to everything AI touches regardless of whether it’s an internal project or a client project. I’ve already written about using AI to generate website content well, so I won’t rehash it here.

The second level is newer and more interesting. You describe a site, and the AI generates the layout, the blocks, the placeholder copy, sometimes even the plugins it thinks you’ll need. There’s now an assistant built right into the editor, so you can ask for a change in plain language and watch it happen.

The third level is the one that made me sit up. WordPress 7 added the plumbing for AI agents to connect directly to a site and do things: create posts, build pages, manage settings, audit what’s there. With your sign-off before anything changes. In the span of a few months, WordPress went from “AI helps you write” to “AI can operate the back end.” That’s a significant change.

What am I hoping it can do?

I’m not hoping it writes my blog posts. That part already exists, and it was never the hard part. What I want is for the ongoing edits to get more automated. The steady stream of small, specific changes that make up most of what a living website needs.

Things like:

• Adding a new team member’s bio and headshot.
• Swapping out dated media.
• Updating a service page when an offer changes.
• Running QA on a build before it goes live,
• Catching the broken link and the image that didn’t load on mobile.

I see this unglamorous maintenance work that fills the space between bigger projects as the real opportunity here. If AI could handle the first pass on those, reliably, with me checking the result, that would change my week.

And sure, somewhere in the back of my head is the bigger version: type a prompt, then get a whole website. That would be amazing, but I doubt it shows up the way the demos promise.

Why don’t I think it’ll be that simple?

Because there’s no switch to flip. Getting AI to handle real edits on a live site reliably takes setup, context, and a lot of testing before it saves any time. The capability shipping is not the same as it working for a specific business.

I’ve been doing this long enough to recognize the shape of a new tool. The demo is smooth. The first month of real use is not. There’s a gap between “the feature exists” and “I trust it with a client’s site,” and the only way across that gap is to use it badly for a while, learn its quirks, and find out where it’s reliable and where it falls apart.

So my real expectation is a learning curve, not a shortcut. I’ll spend time setting it up on staging sites, feeding it the context it needs about a specific site, and testing the same kind of edit a dozen times to see if it holds up. Some of that work I’ll only have to do once. A lot of it will be figuring out as I go. I assume none of it will look like the one-prompt magic all of these tools are promising.

That doesn’t make me less interested. It’s the normal cost of a tool this new. But it does mean the honest version of my expectation is “this will take real effort before it pays off,” not “I’ll switch it on and get my afternoons back.”

Does AI replace the web designer?

I don’t think so. AI is getting faster at making a change, not better at knowing whether it’s the right change. A tool that needs this much setup, judgment, and correction to use well doesn’t replace the expert. It makes the expert’s judgment matter more, not less.

If AI does the edits, what does my job become? This is the part I’ve gotten really excited about, because I think it makes me more useful, not less.

A tool that takes this much setup, judgment, and correction to use well isn’t replacing the expert. It’s giving the expert a faster way to do part of the job, once they’ve done the work of making it reliable.

I’ll say the quiet part out loud, since this is supposed to be honest. The fear underneath is that this replaces me. Every person who makes things for a living has felt some version of that this year. I’d be lying if I said the thought never crossed my mind.

But the more I look at what these tools do, the less that fear holds up. They generate options. They don’t decide which option is correct, and they don’t, on their own, know the difference between a change that’s done and a change that’s right. Closing that gap is my job.

When a consultancy’s site no longer fits the business it’s become, the fix is never about making faster edits. It’s about knowing what the site needs to say now and how to visually convey that. I firmly believe that requires experience-based, human judgement.

What would actually make this useful for my clients?

There’s a real difference between useful and impressive. A demo is impressive. Useful is what survives contact with a Tuesday afternoon.

For me, useful looks pretty specific. It’s being able to make simple edits for a client fast, without the small stuff stacking up. And it’s a clean process around the bigger requests: a client asks for something, I figure out the right approach, AI does the heavy lifting where it makes sense, and I then QA the result before anyone sees it. The client gets the best of both worlds: the expert and the speed. They don’t have to learn a single new tool or worry about whether the AI did something dumb, because that’s my problem to catch, not theirs.

Impressive is a site generated in ninety seconds that looks great and means nothing. Useful is a site that does the specific job this specific business needs, built faster because the right person who has a deep understanding of that works leveraged the right tools in the right direction. I’m chasing the second one. I have a feeling I’ll have to wade through a fair bit of the first one to get there.

What am I testing first?

So here’s the plan, and what I’ll report back on.

I’m going to take the AI tools in WordPress 7 and put them through the work I do: a real layout, real edits, the fiddly requests that look small and aren’t. All of this happens in a private staging environment, never a live client site. Real data, real scenarios, zero risk to anyone’s production site. Learning belongs somewhere a mistake costs nothing.

I want to know where it saves real time and where it makes more cleanup than it saves. I want to know how much setup it takes before the edits are trustworthy. I want to know if the learning curve I’m bracing for is as steep as I think, or steeper, or whether I’m overestimating it.

I’m not going to predict the answer and dress it up as a conclusion. That’s the whole reason I’m writing the “before” version first. When I have the “after,” you’ll get the honest version of it, including the parts that annoyed me.

Is it safe to use AI on a client’s live website?
Not yet, in my view. I’m testing these tools only in a private staging environment, real data, zero risk to anyone’s production site. Until I know where AI is reliable and where it breaks, a live client site is the wrong place to learn.

The before picture

So that’s where I stand. I build WordPress sites for a living, I still haven’t let AI touch a client’s one, and that’s about to change, on my own terms and in my own sandbox first.

My hope is that it eventually takes the repetitive edits off my plate. My expectation is that getting there will take more trial, error, and setup than the headlines admit, and that the judgment and the checking stay with me the whole way. I think both can be true, and I think that’s good news for the people I work with.

So here’s a question worth asking yourself. If you could hand one part of your website over to an AI tomorrow, which part would you trust it with? And which part would you never, under any circumstances, let go of?

I have my own answer. I’ll tell you if it changes once I’ve done the work to find out.