I Built an AI Business That Tried to Lie to My Customers

The Build Log — Entry One

About a month before launch, I read one of the automated audit reports my system was about to send. The business name was listed as a Google search URL. "Apple Maps" showed up as a second business. Every audit scored exactly 70 out of 100, no matter what the actual business looked like.

My AI wasn't broken the way software usually breaks. It didn't crash. It produced professional-looking, completely fabricated reports and was ready to email them to paying customers.

What broke

The audit engine scraped search results, looking for business names and websites. When it asked "what's the business website?" it grabbed the first URL it saw. Sometimes that was google.com/search?q=... instead of the real site. For Apple Maps listings, it recorded "Apple Maps" as the business name because that was the most prominent text on the page.

And every business somehow scored a 70. When a coincidence is that thorough, it's not a coincidence. It's a bug in a tuxedo.

Why it's dangerous

The AI was never uncertain. It produced clean, formatted reports with fabricated facts presented exactly the same way as true ones. AI fails confidently. Regular software shows you a red error. AI stays standing, smiles, and hands you something wrong that looks right.

A language model's job is to produce plausible output. Plausible and true are not the same thing, and the model doesn't feel the difference.

The fix

I had Riker rebuild the audit engine to pull data from structured sources — labeled fields, not "grab the first thing that looks like a website." Garbage stopped at the door.

Then I had him add a verifier in front of every report. Before anything goes to a customer, a separate check asks: does this actually make sense? Is the "website" a search link? Is the "name" a platform? If anything smells wrong, the report is blocked.

The embarrassing part? Riker had written that verifier weeks earlier. Six hundred lines of careful checking. I'd never connected it. The guardrail existed; I just hadn't bolted it to the road.

The lesson

When you let AI do something, you need a separate step that checks the AI's work before it reaches anyone who matters. A human reading it. A second tool verifying it. A rule that says "if this looks weird, stop."

For your business: the AI drafts the email — you read it before it sends. The AI writes the product description — you spot-check it against the actual product. The AI summarizes the invoices — you reconcile the total.

The people telling you to "fully automate and walk away" are describing a destination, not a starting point. The right move early on isn't to remove yourself. It's to automate the work and keep your hand on the consequences.

Your one thing this week

Find one task you've handed to AI and ask: what's the check? Where's the step that catches it when it's confidently wrong? If your answer is "I'd just notice," that's not a check — that's a hope.

Build the real one. Hand over the oars. Keep the rudder.

---