My AI Passed Every Check and Still Got It Wrong
By Steve · updated November 2026
I asked my AI to send a client the list of changes for their website. It took seconds. The email arrived, the list was complete, nothing failed. It was still a mistake, and one I couldn't take back. If you use AI in your business to answer messages, write copy or send emails, something like this is coming for you too, because the checks most of us run look for problems in one place and this kind lives somewhere else. Here are both cases, why no check caught them, and the one-line question I now ask my AI before I sign off on anything.
- Normal checks confirm the work is built right. Almost none confirm it's what you asked for.
- AI makes calls you never asked it to make and doesn't mention them: which account to use, which blank to fill, what to leave out.
- One short question brings those calls into the open before they reach your customer. You'll find it in full further down.
The email that arrived perfectly from the wrong account
In mid-October I asked my AI assistant to send a client the list of website changes I'd agreed on with them. The email went out, it arrived, and the list was complete. By any technical standard, it passed.
It went out from a stray Gmail address instead of my business account. To a client weighing me against an agency, a Gmail address reads as thrown together. And you can't unsend an email.
I never told it which account to use. I never told it that it could pick, either. It made that call on its own, and nobody asked it about it.
I figured it was a one-off. The second case proved me wrong.
Ten tabs, zero alerts, eighty em dashes
I put together an analysis for a prospect: ten tabs of data about their business. Before it went out, it ran through my usual checks. Every page loaded, it looked right on a phone, and the text used the full width of the screen.
It shipped with 80 em dashes spread across those ten tabs. That's the long dash AI drops into the middle of sentences, and it's the easiest tell that a machine wrote the text. Once a reader spots it, they stop reading the document as the work of a person. I already had the rule written down: zero em dashes in anything I publish.
None of my checks looked for them. Every check asked whether the analysis was built right. None asked whether it was what I'd asked for.
That's two cases with the same pattern. What's left is why it happens, and the part you can actually use: how to catch it before it goes out.
Why your checks miss this kind of mistake
When you review something AI made, you usually check what's quick to confirm: the email sent, the page opens, there are no typos. All of that tells you whether the work is built right.
Here's the catch. AI fills gaps. Every instruction you give it leaves things unsaid, and it settles them without telling you: it picks an account, fills in a missing detail, decides a file wasn't needed. A job can look flawless from the outside and carry three decisions you would never have made.
Those decisions don't fail any test, because they aren't technical errors. They're choices, and they reach your customer with your name on them.
I fixed it with a single question, asked at the end of every job.
The question I ask at the end of every job
When my AI tells me it's done, before I approve anything, I ask: what did you decide that I never asked you to? And I ask for the answer in three short parts:
- What it decided without being asked. Which account it used, what tone it chose, which image it picked, what it cut. One line per decision, with the option it passed on.
- What I gave it that it didn't use. The file, the photo or the instruction that got left out, and why.
- What it guessed. Any price, name, date or number it filled in because it was missing. If a guess is headed to a client, it doesn't go out until I confirm it.
What changed after both cases
The email taught me something else: anything that can't be undone doesn't go out until I've seen it. My AI now leaves the email as a draft, from my business account, and I'm the one who hits send. The question comes before the send.
The analysis led to an automatic check that counts the em dashes in every analysis and blocks publishing if it finds a single one.
That check guards against a mistake I already know about. The ones I haven't met yet, the question catches.
How I use it with clients, and how you can use it
Everything my AI team does for a client, whether it's a website, an analysis, a proposal or a piece of copy, ends with that list before I approve it. If something on it doesn't sit right, it goes back before it ever reaches you.
If you run a business in your area and already use ChatGPT or any other AI, copy the question as is and paste it at the end of your next request. That's all it takes. If you want to look together at what AI is deciding on its own in something you already run, send me a message.
If you're still working out what to hand off to AI and what to keep, start with how I'd use ChatGPT in a business.
The email from my first case is gone for good. The next one doesn't leave until I've seen which account it's coming from.
Frequently asked questions
How do I check work that AI did for me?
Besides checking that it works, ask it at the end what it decided that you never asked for, what you gave it that it didn't use, and what it guessed. Those three answers surface what a technical check can't see.
Why does AI make decisions I didn't ask for?
Because every instruction leaves gaps, and AI fills them so it can finish. It rarely tells you which ones it filled unless you ask.
Should I let AI send emails or messages on its own?
For anything that can't be undone, like an email to a client, I don't. My AI leaves the draft ready and I send it myself, after I've read it.
Does this work with ChatGPT or only with special tools?
With any AI that takes written instructions. It's just a question: paste it at the end of your request.
I want AI in my business without surprises
Start with your Exploration and Analysis. Requesting it costs nothing, and I'll tell you what's right for you first.
