By Eric Mahler Founder and Managing Partner, Aretos Advisory
Every vendor selling AI tools into the enterprise right now is, in one form or another, making the same promise: this system will help your people make better decisions. Faster insights, sharper recommendations, fewer blind spots. It’s a reasonable pitch, and in a lot of narrow use cases, it’s true.
But there’s a category error hiding underneath most of these conversations, and it’s costing organizations more than the AI budget line item suggests. Judgment gets talked about as if it’s a feature you can add to a system, something that ships in the next release, gets refined with more training data, eventually reaches parity with a human decision-maker. It isn’t a feature. It’s the thing that decides which features matter, when to trust an output and when to override it, and what to do when the system is confidently wrong.
You cannot automate your way out of needing that. You can only decide whether you’re building it deliberately or losing it by default.
What actually happens when judgment gets outsourced quietly
Nobody sits down and decides, on purpose, to stop exercising judgment. It erodes gradually, through a specific and predictable pattern: a system produces a recommendation, the recommendation is right often enough that people stop checking it, and eventually the checking stops entirely, not because anyone decided the system was infallible, but because the habit of questioning it simply atrophied.
This is where a lot of AI deployment conversations go wrong. Leadership asks whether the tool is accurate. That’s the wrong first question. The better one is: what happens the first time it’s confidently, plausibly wrong, and does anyone in the process still have the standing, the information, and the instinct to catch it?
Accuracy is a property of the system. Judgment is a property of the organization around it. You can have a highly accurate system embedded inside an organization that has completely lost the capacity to question it and that combination is more dangerous than a mediocre system everyone still checks.
The trade organizations don’t realize they’re making
Every time a decision moves from a person to a system, an organization is trading something specific: the slower, more expensive, more visibly fallible process of human judgment, for the faster, cheaper, more consistent output of a model. That’s often a good trade. It’s rarely evaluated honestly.
What tends to disappear in that trade isn’t just speed. It’s the specific kind of accountability that comes from a person having to stand behind a decision, explain their reasoning, and live with being wrong in front of colleagues who watched them make the call. A system doesn’t feel the weight of a bad outcome. It doesn’t hesitate before a decision that’s technically defensible but morally uncomfortable. Those aren’t bugs to be engineered away. They’re exactly the parts of judgment that made it worth having in the first place.
Organizations that deploy AI well tend to ask, explicitly, what they’re willing to let the system decide alone, what always requires a person in the loop, and what requires a specific person, someone with the authority and the incentive to actually push back, not just someone technically positioned to click approve.
What this looks like in practice
A few markers separate organizations that are managing this well from organizations that are drifting into it by accident:
They can name, specifically, which decisions the system makes alone versus which ones require human sign-off and that list isn’t just “everything above a certain dollar threshold.” It reflects where consequences are hard to reverse, not just where they’re expensive.
They’ve built in friction on purpose, in places where friction is uncomfortable but necessary. A system that makes it easy to accept a recommendation and hard to meaningfully question one is training people toward passivity, whether that was the intent or not.
They treat a string of correct outputs as a reason for more scrutiny, not less. The moment a system has been right often enough that nobody’s checking anymore is the exact moment the check matters most.
They know who is actually accountable when the system is wrong, and that person knows it too, not as an abstract line in a governance document, but as something they’d be able to explain plainly if asked.
The real question
The pitch every AI vendor makes is some version of “this will help you make better decisions.” The question worth asking back isn’t whether that’s true. It’s what happens to your organization’s capacity for judgment if you accept the help without keeping the muscle exercised. Judgment doesn’t fail all at once. It fails quietly, one skipped check at a time, until the day it’s needed and nobody remembers how to do it.
About the Author
Eric Mahler is the founder and Managing Partner of Aretos Advisory. He has spent thirty years advising boards and executive teams across industries. He is the author of The Center of the Compass, forthcoming from Amplify Publishing in December 2026, and a keynote speaker on leadership, AI governance, and organizational resilience.
ericmahler.com | Connect on LinkedIn
Aretos Advisory works with boards and executive teams navigating AI transformation, governance, and leadership coherence. If this piece reflects challenges your organization is working through, we’d welcome the conversation.
© 2026 Aretos Advisory. All rights reserved.