The viewer's win that felt a bit too easy
During our August 27, 2026 Office Hours livestream, an agency owner shared a story that might sound familiar to anyone easing into AI-assisted work. In her own words:
I was able to successfully tackle projects that I would never have dared to start six months ago, including removing malware from a website and building a small plugin. But my question is, how trustworthy is AI in this kind of situation? I realized it could have instructed me to add or remove something that could make things worse. It all turned out well, but I was being naive to venture into this with only Claude as my guide.
And here's the part worth paying attention to: she'd done all of this without connecting Claude directly to the website or using any agentic tools.
All she did was describe the problem, share snippets of code, and ask questions back and forth, following the AI's instructions closely to get each of her projects done. It worked out for her, but she couldn't shake the feeling she'd gotten lucky with the accuracy of the AI's help.
That's a perfectly valid concern, considering AI tools can still get things wrong or invent details with total confidence. Also, a couple of questions instantly come to mind: "What if the AI had told me to remove the wrong file?" and "What if what I am doing makes the malware worse?"
The question that most bothered her, though, was if she had been too naive to lean this much on an AI assistant with no other safety net.
Start with guidance, build confidence
Nathan Ingram, the livestream host and agency trainer here at hosting.com, had a response that reframed the situation into something more positive.
Rather than treating the hesitation our viewer felt as a red flag, he pointed out that it's exactly how most agency owners who are now fully comfortable with agentic AI started. As Nathan put it:
No, I would not say you were being naive at all. I would say welcome to the club. Once you see what AI can unlock for you, it's, what can I do next?
He described a pattern that, as we mentioned earlier, is familiar across the Agency Success community. You start by just letting the AI tool guide you, while you do all the manual work, as you get comfortable with the process. Then, you experiment with agentic AI, get comfortable with that process as well, and little by little start handing over the execution to it.
Finishing a project after being guided by AI can be the first step to learning how to do something wholly new; it's not a shortcut that avoids learning.
The safety net that matters
None of what we've talked about so far means you can throw caution out the window, though. Nathan made it very clear, and we are sure you know this already, that AI-assisted work can break things in spectacular ways. Countless articles online about AI destroying a database or bricking a site are a testament to that.
So, the first thing you should always do before you start experimenting is ensure you have a way to revert back to a version of your project prior to working with the AI. In other words: have a backup of your project. Regardless of how much you trust the process, always have a backup.
But that's not the only safety net you could use in your workflow. Here are a few other suggestions:
As you can see, only our last suggestion asks you to slow down a little as you go. That trade-off is worth it: if something does go sideways, you have a quick way to undo it, turning what could be a crisis into a minor inconvenience.
How much should you let AI do?
One thing that is often misunderstood by people new to AI workflows is that it's not a single, set-and-forget setting.
It's a spectrum that depends on each individual user and project that spans AI simply advising you and letting it execute changes directly, only stopping when it hits a genuine decision point. Nathan described it very well during the livestream:
You can choose with Claude what level of interactivity you want, whether it stops and asks your permission for everything, or whether it just kind of runs along until it hits some sort of barrier.
While he's talking about Claude, this is a concept that can be applied to all major AI agents. This idea, often called an agent's level of autonomy, is something researchers and AI companies think carefully about when designing these tools. Where you land on the spectrum of an agent's level of autonomy is a personal call.
It's also fine if your preferences shift over time. Someone earlier in their AI journey might want their AI assistant to check in at every step. On the other hand, someone more experienced might feel comfortable letting a tool handle more of the execution directly.
Where to go from here?
There is a practical reason to think about moving forward in the spectrum as you get comfortable: speed. Manually reading AI-generated instructions and doing the work yourself adds a lot of friction to a project.
Nathan estimated a malware cleanup that takes hours could take a fraction of the time if the agency owner only stepped in when necessary and let the AI do the rest of the work.
Despite that, though, if you are still in the "advise me, I'll do the rest of the work" phase of working with AI, you shouldn't rush it. It's the natural starting point and it's where you'll learn the most.
Just make sure you have a good backup (several if you can), keep an eye on and stay engaged with what the AI is suggesting, and push back when you don't like something. Your confidence in the process will build gradually, after which you'll see how you are speeding up.
And, in case you have other questions about AI workflows, have an AI first win you'd like to share, or want to learn more about AI in general, WordPress, or hosting, join us for Office Hours, every Thursday at 2 p.m. ET. We'd love to have you.
FAQ
What's a good way to double-check AI-suggested code before running it, even with a backup in place?
Beyond having a backup, it helps to read through what the AI is suggesting line by line before applying it, especially anything that deletes files or writes directly to the database. If you're using version control like Git, committing your current state first gives you an easy way to see exactly what changed and roll back a single file instead of restoring an entire backup.
Yes. Chatting with an AI keeps a human in the loop for every single action, since you're the one copying, pasting, and applying each change yourself. Connecting AI directly to your site through an agentic tool removes that manual step and lets the AI make changes on its own. Both sit on the same autonomy spectrum, but starting with the conversational version, like our viewer did, is a common and reasonable way to build trust before moving to something more hands-off.
Are there website tasks that are too risky to hand over to AI, even once you're comfortable with the process?
Yes. Tasks with a large blast radius if something goes wrong, like building a custom payment integration, generally aren't good candidates for full AI autonomy, no matter how much experience you've built up elsewhere. For those, keeping a human reviewing every step, or skipping letting the AI do it all entirely, is usually the safer call.