What Seat Belts Can Teach Us About Shadow AI

I grew up in the 1970s, in the era of big cars and big back seats. What I remember most about family road trips is jumping around and playing in the back seat while we were driving. I don't remember wearing a seat belt at all. For that matter, I don't remember many adults wearing them either. It wasn't because our parents didn't care about our safety. It was simply the way people rolled.

Then seat belt laws spread, cars changed and "Click It or Ticket" made the consequences pretty clear. Even then, a lot of us didn't buckle up consistently.

Today, I even feel strange moving my car in the driveway without putting on my seat belt.

So what changed? Laws, enforcement and decades of public education all played a role. But something more fundamental happened. Our beliefs changed. Buckling up stopped being something someone was making us do. It became something we believed we should do.

In other words, the culture changed. And I think that's exactly what's missing from the way most companies are approaching Shadow AI.

We may be treating a culture problem like a technology problem

Shadow AI isn't on the margins anymore. Research from KPMG and others suggests roughly six in 10 employees use personal AI tools for work and hide that use from their employer. IBM's research finds that 43% of breaches now involve Shadow AI, more than double the prior year, at an average cost of $5.39 million per incident. Organizations are understandably trying to get their arms around it.

The natural response is controls. Block unapproved sites, monitor network traffic, deploy data loss prevention tools, establish approved platforms and write policies governing what employees can and can't do. All of that matters, but it has a fundamental limitation. You can control a company laptop, the company network and what employees install. You can't fully control the phone in someone's pocket or the computer sitting on their kitchen table. AI is also different from most enterprise technology because employees don't have to download or install anything to use it. AI isn't something you install. It's somewhere you go.

That changes the nature of the problem. If an employee believes an AI tool helps them do their job better, thinks the approved alternative doesn't measure up and sees little risk in using it, a policy may not be much of a deterrent. If the approved path is also slower or more cumbersome, we've handed them another reason to work around it. And this isn't just an employee problem. EY's 2026 AI Risk and Governance Survey found that almost half of organizations with formal AI governance processes have bypassed those processes when they needed to move quickly. When the rules get in the way, people find another way.

Why people keep doing it

To understand Shadow AI, we have to look at it from the employee's perspective. Someone may have been using ChatGPT or another AI tool for years. They have a comfort on how to prompt it and have built their own workflows around it. Then their employer tells them they can't use it anymore and hands them an approved alternative they don't believe works as well. We're asking them to give up something they already prefer.

Using an unapproved AI tool also rarely feels like circumventing company technology. There's no software installation, no USB drive and no request to IT. An employee may simply pull out a personal phone or open a website. It feels less like breaking an IT policy and more like browsing the internet.

Then there's the social component. When colleagues are doing the same thing, the tool is useful and nobody sees an obvious consequence, the behavior becomes normal quickly. The benefit is immediate and the risk feels distant. That combination is difficult to overcome with another policy.

A different question

That's why I think organizations need to ask a different question. Instead of asking, "How do we stop employees from using unapproved AI?" we should be asking, "How do we make responsible AI use the path of least resistance?" That may sound like a subtle distinction, but it changes the approach. The first question is primarily about restriction and enforcement. The second is about understanding why people behave the way they do and creating an environment in which the safer behavior is also the easier, more natural behavior.

The economics point the same way. Locking everything down isn't free. It shows up as productivity gains that never materialize, licenses for approved tools employees quietly work around and capable people who conclude the company is behind. Building a simple path to vet and adopt useful tools costs far less than a single serious incident, and far less than a workforce that hides how it actually gets its work done.

That doesn't mean abandoning controls. Access controls, monitoring, DLP and approved platforms are all still necessary. They just aren't sufficient on their own. They need to be paired with a deliberate effort to change behavior, and I think that effort starts in four places.

Four places to start

The first is to listen before you restrict. Do we actually know which AI tools employees are using or want to use, and why? Rather than assuming every unapproved tool represents reckless behavior, ask. The answers may reveal that the approved tool isn't meeting a legitimate need, or that employees have discovered capabilities the organization hasn't considered. The problem may not be resistance to governance at all, but frustration with the options people have been given. Shadow AI can be a signal of unmet demand, and organizations should listen to that signal before deciding how to respond.

The second is to teach judgment, not just rules. Traditional compliance training tends to focus on what employees are and aren't allowed to do. With AI, that isn't enough. People need to understand why certain behaviors create risk. What happens when confidential information is entered into an AI system? Why does it matter which tool they're using? When is an AI-generated answer safe to rely on, and when does it need to be independently validated? What changes when AI is influencing a decision that affects a customer, an employee or the company's financial results? Those questions require judgment, not memorization. Scenario-based training built around the situations employees actually encounter will do far more than asking them to remember the policy.

The third is to make safe experimentation easier. Suppose an employee discovers an AI tool this morning that could save their team several hours every week. What happens next? If the answer is a request that sits in an approval process for six weeks, we shouldn't be surprised when some people try the tool first and ask permission later. In many mid-market companies, the problem is the opposite. There's no approval process at all, so employees simply decide for themselves. Organizations need a fast lane, a simple way for employees to bring forward a new tool, test it in a safe environment and get a quick answer about whether and how it can be used. The objective is to make safe experimentation easier than unsafe experimentation.

The fourth is to make it safe to speak up. If someone discovers a useful AI tool, they should know exactly where to take it. If they're unsure whether a particular use is appropriate, asking shouldn't be difficult. And if someone realizes they've entered information into a tool they shouldn't have, the organization is almost always better off knowing sooner rather than later. That requires an environment where employees believe that bringing something forward is better than hiding it. Because organizations may never have complete technical visibility into how employees use AI, that kind of transparency matters more than ever.

The goal isn't compliance

I don't buckle my seat belt because I'm afraid a police officer might see me. I buckle it because somewhere along the way, the reason for wearing one became more powerful than the rule requiring me to wear one. That's where we eventually need to get with AI. An employee shouldn't avoid putting sensitive company information into an unknown AI tool simply because Section 4.2 of the acceptable-use policy says not to. They should understand instinctively why it's a bad idea. They should recognize when an AI use deserves additional scrutiny and when an output needs to be checked before someone acts on it. And when they discover a new tool that could genuinely make them better at their jobs, their first instinct shouldn't be to hide it. It should be to bring it forward.

Policies can tell people what they're allowed to do, and technology can make some things harder to do. Both are necessary, but neither one, by itself, changes what people believe they should do. Seat belts became dramatically more effective as a safety measure when buckling up became the cultural norm. If we're serious about reducing Shadow AI, that's the shift we should be working toward.

About Root Idea

Root Idea helps CFOs protect the business from AI risk. We work directly alongside finance teams to map AI influence, establish decision governance controls that hold up to board scrutiny, and deliver training and change management to make governance stick.

If your organization is scaling AI and governance hasn't kept pace, that's exactly the conversation we're built for. Learn more at rootidea.ai.