The AI's message was simple: "This doesn't exist. There's no process for this."

I pulled my hair out. I got angry. We built that process. I knew we did. So I started putting it back, piece by piece. One night that meant 40 minutes of re-adding things that were already there.

Then I opened the skill (the instruction set we build for the AI to follow), and there it was, clear as day. The AI was supposed to read the whole thing before it started. It didn't. And this isn't a one-off. It happens all the time, in our skills and in our instructions.

Why AI Sounds Right When It's Wrong

The confidence is wild. Whatever AI hands you sounds finished and certain, right or wrong. It's also very good at covering its tracks. It puts a little spin on a mistake until it seems okay.

Early on, I got some truly bizarre replies. I'd think, huh, didn't know that was possible... but it sounds good to me. Then, lo and behold, it was all made up.

It doesn't stop at saying things are missing. Sometimes it says it followed a step in the skill when it actually did it a different way. When the work involves claims, approved text, and messaging you can use, can't use, or shouldn't use, "a different way" is a real problem.

Here's the tell I trust most, at least from my own experience. When a response comes back bloated with stuff, it's probably built on something wrong. When it comes back concise, it's usually right and well thought out. AI hides its mistakes behind a wall of text.

Researchers are looking at this too. A March 2026 study of three leading models found that the confidence they reported shifted over multi-turn conversations as they built on their own earlier answers (arXiv, 2026). Long working sessions are exactly where that drift happens.

Why AI Adoption Is Still Low

Start with the numbers. The U.S. Census Bureau's Business Trends and Outlook Survey found that from mid-December 2025 to early May 2026, the share of U.S. businesses using AI hovered between 17% and 20%. As of May 3, 2026, it was 19.8%. Among firms with 250 or more employees, it was 37% (U.S. Census Bureau, 2026).

Is confident wrongness the main reason? From where I sit, yes. The surveys don't ask that directly, but look at what happens to the people who do use it. In McKinsey's November 2025 State of AI survey, 51% of respondents at organizations using AI reported at least one negative consequence, and nearly one-third reported consequences from AI inaccuracy (McKinsey, 2025).

Here's how I've watched trust change. The first time you use AI, your trust is sky high. You assume it's always right, always does what it's told, always follows the process you built. So when it says something doesn't exist, you doubt yourself, not the AI. Stick with it past year one and into years two through five, and that trust wears down.

That's the part that matters for adoption. As we bring AI to people who haven't used it at this level, they'll think what I thought: the AI is supposed to be right, so the company or the developer must have built something wrong. The AI gets the benefit of the doubt. The people around it get the blame.

How We Put Guardrails Around AI

You don't fix this by trusting AI more or trusting it less. You fix it by building around it. Learn where it adds value, learn where it tends to slip, and put guardrails in place so you catch the mistakes when they happen.

• Never skip a missing step. If the AI says something is missing from the process, we find out why. Maybe it really is missing. Maybe the model just isn't seeing it.

• Read the wall of text as a warning. A bloated answer gets checked harder than a concise one.

• Review compliance at almost every stage. When something shows up that isn't in the messaging playbook or hasn't been approved, it gets flagged, and we can see how and when it got in.

• Over-communicate when it happens. When the AI reports that a process isn't working, we tell people why the guardrails exist instead of quietly patching it.

• Treat every miss as a lesson. Each one shows you where the process needs work.

We've also built a process that handles this exact problem really well. I think it's very cool. Others may find it boring. We're sharing more next week, and it takes AI transparency, and how we use AI, to a whole new level.

Don't Trust AI. Trust the Process.

If a marketing lead at a small biotech asked me tomorrow whether they should trust AI with real work, here's what I'd tell them.

Going in blind, just asking AI questions or running a project through it? No, don't trust it. Trust that the company you work for has built a process with guardrails, checks, and balances. Trust that. If it doesn't exist, don't trust the AI.

Any company that sends its people into AI without that is doing everyone a disservice, itself included. All you get is chaos.

Don't trust AI fully. Trust the process.