Every organisation writes policy assuming people will follow it. Most people don't, and it's rarely carelessness. The trade-off they're actually making usually doesn't match the one the policy assumed they'd make. This conversation gets into why that gap exists, what it looks like specifically with AI tools, and what actually closes it.
The Benefit Is Personal. The Risk Isn't.
People run a cost-benefit trade every time they pick up a tool they weren't given, and from the outside it can look irrational. It isn't. An employee feels the productivity gain from an AI tool the moment they use it. The risk that tool creates rarely touches them at all. It sits with the organisation, and it tends to surface much later, if it surfaces at all.
"If I'm the employee, and the tool helps me improve my own productivity, that's something I feel immediately. But the risk is disintermediated, it's to the organisation. It doesn't affect me directly." - Dr. David DeFranza
There's a useful comparison here to climate change, where the same mismatch shows up: the cost of changing behaviour is personal and immediate, the benefit is collective and years away, so behaviour barely moves. Training people on policy doesn't fix this on its own, because the training doesn't touch who holds the cost and who holds the benefit. What does move things is incentive design, making the approved path pay off in roughly the same currency the unapproved one already does.
A 90-10 Split, and Why It Feels Different with a Machine
A decades-old economics experiment makes the pattern concrete. Two players split a sum of money. One proposes how to divide it, the other can accept or reject the offer, and a rejection means both players walk away with nothing. Play this with two humans and people reject uneven splits constantly, even past the point where rejecting costs them money they'd otherwise keep.
Replace one player with an AI and the rejections drop off. People start accepting splits that are far more lopsided, closer to the outcome a purely rational calculator would choose. Trust has surprisingly little to do with it. What changes is the belief that AI is a rational actor rather than an emotional one, a belief most people picked up from decades of film and television long before they ever opened a chatbot. The same instinct likely shows up at work, in how little scrutiny an AI tool gets compared to what a new hire or a new vendor would face, simply because handing something to a machine doesn't feel the same as handing it to a person.
Where AI Sharpens Judgment, and Where It Gets in the Way
That shift has a practical flip side inside a business. For work that rewards consistency and calculation, teaming up with an AI tool tends to make judgment sharper. Decisions that lean on emotion, intuition, or reading a room are a different story. There, the same pull toward a more "rational" mode of thinking can quietly work against the person making the call.
Knowing which kind of decision is on the table matters before deciding whether AI belongs in it at all.
What Actually Changes Adoption
Tighter policing of who's using what tool rarely works on its own. What tends to work is alignment.
"We need to think about how we align the incentive people get from productivity improvement with the incentive to be compliant with a corporate risk management policy." - Dr. David DeFranza
People also learn a tool's real limits fastest through direct, low-stakes experimentation. A policy document can tell them what to expect. Using the tool and watching it fail at something teaches them faster.
The risk was never really about the tool. It's about knowing who's carrying the cost of using it, and why the incentive to look away is usually stronger than the incentive to raise a hand.
