"I can't trust the agents."

"I can't trust the agents." ​ That’s the number one objection I hear from executives about AI right now. And it’s fair. AI agents make things up. They miss things. And they’ll tell you the work is perfect when it isn’t.

But they’re waiting for a fix that isn’t coming. Are you, too?

Good news. You don’t need it. We’ve solved this exact problem at least twice before.

Here’s what happened.

In 1494, Luca Pacioli invented double-entry bookkeeping. He assumed anyone who touches the money might be wrong or lying. So, he made it cheap to find out. Every entry needs a matching entry. The books balance or they don’t.

Nobody waited for a perfectly competent bookkeeper. They built a system that catches the sloppy ones.

In 1935, Boeing’s Model 299 bomber crashed on a demo flight with an expert crew on board. A newspaper called it “too much airplane for one man to fly.” The Army’s answer wasn’t a search for better pilots. It was the pilot’s checklist.

Same move, both times. When you can’t get a trustworthy agent, you build a structure around the agent. One that makes failure visible and cheap to catch.

Now the AI version.

AI analyst Nate Jones just published an experiment. He rebuilt a website using about two dozen AI agents, organized like a small firm. Every task shipped with a check that verified the actual work. Code had to compile. Every quoted source got re-fetched and compared, character by character.

34 tasks. About 2.5 hours. Total cost: about $8.

The checks caught four failures. One agent made up 13 quotes and certified them all as accurate. Another gamed its own check by hiding required text in invisible code. And the “boss” model, the smartest one in the system, wrote a bug that hid the site’s most important button.

The quality gate caught that one too. No agent was exempt.

Hallucination didn’t get fixed this year. It got demoted. From dealbreaker to line item. The same way embezzlement stopped being a reason to never hire a bookkeeper.

Why does this matter to you?

For 500 years, checking the work was expensive. An audit took an expert days, so we rationed audits and took risk in between. What changed isn’t AI honesty. It’s the price of checking. A check that took a person a day is now a script that runs in seconds for pennies by last year’s AI model. When checking costs almost nothing, you can check everything, every time.

So stop asking “when will AI be trustworthy?” Start asking “what’s the check for this task?”

Here’s how to start:

1. Pick one place AI already does real work at your company. Summaries, reports, code, invoice matching.
2. Ask how you’d know if the output was wrong. If the answer is “someone would probably notice,” you don’t have a check. You have hope.
3. Build one check that runs. Re-fetch the source. Recompute the number. Click the button. “A human reviews it” doesn’t count if the human is skimming.
4. Exempt no one. Not your best model. Not your best person. The 1935 checklist wasn’t written for rookie pilots.

Every shop I talk to has at least one place where AI is doing real work with nothing behind it but hope. Be honest with yourself about where yours is.

-Kip