The Best Performers Bend the Rules – Should AI Be Afforded the Same?

The Best Performers Bend the Rules - Hero Image

One of the top performing employees I worked with during my time in HR was constantly late. They generally managed to get their shit together for big important meetings but otherwise was very rarely on time to work or any of their less important obligations.

I wondered how they got to skate on being late to everything until I found out that this person was great at their job and was usually late because they were doing something that was above and beyond (yet very necessary).

This wasn’t an issue until a year later when an employee we were placing on a PIP for, among other things, being late brought it up. Why was this an issue only for them, they wondered?

That lack of consistent application came rushing back as I read some new research on this from GCheck called, rightfully so, The Performance Privilege.

It found that around half of managers backed a policy exception for a top performer versus just above a quarter for the typical one. More than 80% of managers say that their policies should apply equally but nearly 60% say that top performers get more leniency.

If that seems unfair, you’re right. If that seems like human nature, you’re also right.

Everyone seemed willing to forgive the actions of someone who routinely overdelivered, even if their actions were unprofessional and often wasted valuable time. They were more hesitant when those same actions came from someone with a less stellar track record of success.

Where does AI fit in?

New research from Mercor shows that frontier models regularly miss risk in routine enterprise work when those risks aren’t explicitly laid out. These are the types of tasks that I hear about getting handed to models. In one example, a frontier model deleted a file that should’ve been preserved under legal discovery. Even though it had the context of the task (including a document that mentioned what types of files should be preserved), the risks and boundaries were not explicitly stated when assigning the task.

AI is great at a ton of tasks. When given proper instructions, context and model selection, it can go faster, see blind spots, and be more consistent than a human can ever hope. In short, it is a high performer in many of its most critical outputs.

But it also doesn’t always behave.

I don’t anthropomorphize AI the way some folks do but it is clear that we must acknowledge the human nature of looking past defects in these tools when they are otherwise providing value. And for vendors, it’s critical to think about what controls you have in place that make risks explicit if you are depending on their models for answers.

Good luck?

Share your comments: