First teach it. Then test it. Then trust it.
EXPERT OPINION BY ENTREPRENEURS’ ORGANIZATION @ENTREPRENEURORG
This article was an EO Netherlands member who is a coach, keynote speaker, and the founder of AIPO Network. Van der Zwart shares why he is implementing a more rigorous testing procedure on his AI agents before deploying them
Amid reports that hundreds of AI agents went rogue in the OpenAI-Hugging Face hack, I am doubling down on training agents more thoroughly before authorizing them to act.
Recently, my AI operations manager failed its exam. The most important hire in the whole setup — the one that coordinates every other agent — sat for the test I built to gate deployment and didn’t pass. Not spectacularly. Just clearly enough that putting it to work would have been a mistake.
It went back to school
So, it went back to school. I gave it more curriculum, more worked examples, and another sit-down with the same benchmark. The second time, it passed.
Refreshed leadership advice from CEO Stephanie Mehta
An Inc.com Featured Presentation
Still, I found something oddly reassuring in its failure. A test you can’t fail isn’t a test. The role I most need to trust is exactly the one where I want the bar to bite.
That small moment is the closest thing to a thesis for a question I’ve been chasing: Can you run a business at which 80 percent of the work is done by AI and 20 percent by people? Not as a slide at a conference, but as real work, for real clients, where the output either holds up or it doesn’t.
I’m still in the middle of finding out. What follows is a dispatch from inside the experiment, while things are still messy enough to be honest about.
Source: www.inc.com
