AI Can Do It. But Do You Know What It’s Doing?

Somebody called us last spring with a site that had stopped taking orders. Not down — the site loaded fine, looked fine, and quietly failed at checkout for a chunk of visitors.

They had made a change a couple of weeks earlier. They had asked an AI assistant how to do it, followed the instructions, and it had worked. The page did the thing they wanted. What the instructions did not mention was a second setting that also needed to change, and the failure showed up only on certain browsers, which is why they did not notice for two weeks.

The advice was not wrong, exactly. It was incomplete in a way that only somebody who had done that job before would catch. And that is the pattern I keep running into, so I want to talk about it plainly.

The problem is not that it is wrong. It is that it is 90 percent right.

If AI produced obvious garbage, nobody would have a problem. You would look at it, see it was nonsense, and throw it away. That is easy.

What it actually produces is confident, well-organized, mostly-correct output. The structure is right. The tone is right. Eight of the nine steps are right. And the one that is wrong looks exactly like the eight that are not.

To catch it, you have to already know the answer. That is the whole problem in one sentence. Verification requires the expertise that automation was supposed to replace.

What this looks like in the real world

These are all things I have personally cleaned up in the last year or so. Details changed, but the shape is accurate.

The code that worked and should not have

A form that saved submissions correctly and also stored them somewhere a stranger could read them. Nothing in the behavior of the site would tell you. The site worked. It worked all the way up until it was a problem.

The plugin update that was “safe”

An assistant confirmed a plugin update was compatible. It was — with the current version of WordPress. It was not compatible with the specific older theme underneath it, which nobody asked about because nobody knew to ask. The site fell over on a Friday.

The policy page nobody read

A business generated a returns policy for their site. It was well written. It also promised a 30-day window on a product line where they had always done 14. That page is a representation to a customer. They found out when somebody held them to it.

The content that quietly lied

A service page that described capabilities the business did not have. Not maliciously — the draft filled in what a business like that usually offers, and nobody edited it down to what this business actually does. They spent three months getting calls for work they could not do.

Why none of these are the tool’s fault

I want to be clear, because I am not anti-AI. I use it every day and my team uses it every day. It has made us faster at real work.

But it does not know your situation. It knows the general shape of the answer. It does not know that this particular site is running a custom theme from 2019, that this client’s payment processor has a quirk, that the last three times somebody changed this setting it broke something two pages away. It cannot know that, because that knowledge lives in someone’s head and in six years of history with a specific business.

It also will not tell you when it is out of its depth. It does not have a face that goes blank. It answers the question you asked in the tone of somebody who is sure, every single time, including the times it should not be.

This is exactly why we are still here

I hear a version of this a lot: if AI can build a website, why do I need Techno Goober?

Fair question. Here is my honest answer.

AI can produce a website. It cannot tell you whether the website is right. It cannot tell you that your booking flow has one step too many, or that the thing you put in your headline is not actually why people hire you, or that the plugin you added last month is the reason your site takes six seconds to load on a phone.

Our job used to be significantly about production — making the thing. Production is genuinely cheaper now, and we have passed a lot of that on. Our job today is more about judgment: deciding what should be built, catching what is quietly wrong, and being accountable when something breaks at 4:45 on a Friday.

That work got more valuable, not less. When output is cheap and abundant, the scarce thing is knowing which output is any good.

How we actually work with it

Nothing exotic. Three rules that have held up:

• Nothing reaches a client site without a human who understands it. Not who reviewed it — who understands it. Those are different, and the difference shows up under pressure.
• Anything that touches money, data, or a promise to a customer gets checked twice. Checkout, forms, policy language, anything with a price on it. The cost of being wrong there is not proportional to the size of the change.
• If we cannot explain why it works, we do not ship it. “It worked when I tried it” is not an explanation. It is a coincidence you have not investigated yet.

What to ask whoever is doing your work

Whether that is us or somebody else, these questions are worth asking, and the answers tell you a lot:

• Who reviewed this before it went live, and what did they check?
• If this breaks in three months, who fixes it, and do they understand how it was built?
• Can you explain to me, in plain language, what this change actually does?

A vendor who can answer those has a process. A vendor who gets vague is handing you output they have not verified, and you are the one who finds out.

The short version

AI can do it. That is not in question anymore. Whether anyone in the room knows what it did — that is the question, and it is the one that decides whether you have a website or a time bomb with nice fonts.

If you have something running that nobody on your team fully understands, that is worth a conversation. We are happy to look at it, and we will tell you if it is fine.