The Risk Words: Guardrails, Shadow AI, and Why “Is AI Approved Here” Is the Wrong Question

AI is what’s called a non-deterministic solution, and that’s a bigger deal than it sounds. From an audit standpoint, you want a process that gives you the same answer every time you run it. The underlying model behind an AI tool won’t do that. Which means the job isn’t to eliminate that unpredictability — it’s to build controls and governance around it so the output at the end of the chain is something you can rely on, even if the model underneath isn’t deterministic on its own.

Guardrails are the technical and procedural controls that constrain what an AI system is allowed to do. This is genuinely one of the hardest parts of any rollout: build guardrails too loose, and people get themselves in trouble. Build them too tight, and the tool becomes useless enough that people quietly work around it — which is exactly how shadow AI gets created. The specific questions worth asking before you roll anything out: What can it do by default? What needs permission? Can people build their own integrations, or only use company-approved ones? And critically — what happens when someone breaks the guardrail? I’ve seen companies with well-written AI policies that nobody enforces, and once people notice nothing happens when the policy gets ignored, that’s its own signal, and not a good one.

Shadow AI is what happens when guardrails are too restrictive or people simply weren’t given what they need. It’s not usually malicious. It’s someone who needs a meeting transcribed and doesn’t have an approved tool for it, so they turn on the AI feature in their personal note-taking app instead. This has been going on since the very early ChatGPT days, and it’s become normalized enough that a surprising number of people use AI tools knowingly outside policy and simply don’t mention it. There was a real case where a law firm reportedly let staff go after client data ended up flowing through a personal AI feature in Grammarly — that’s the kind of exposure this creates. And AI systems tend to be goal-seeking: given a task, a model will often find the fastest technical path to solve it, even if that path routes around the process you’d actually want followed. I’ve heard of someone asking an AI assistant for help sending event emails, and instead of pointing them to the comms team, the assistant offered to help them script a batch send directly through the company’s Microsoft backend via API. Technically, that would have worked. It also breaks about every rule you’d want followed.

AI governance is the overarching structure meant to catch both of these — policy, ownership, training, and oversight. The single biggest miss in most rollouts is ownership. Early on, everyone wants to own AI, which in practice means nobody actually does — it becomes a committee, and committees are slow to hold anyone accountable. Real governance means one named person with actual authority to make calls, alongside required (not optional) training that actually gets updated as the landscape changes, and — maybe most importantly — a real path for people to surface a new use case or request an exception. If people don’t have a fast, known way to ask “can I use this for that,” they’ll just use it anyway, and now you have shadow AI again.

The question that actually finds shadow AI isn’t “is AI approved here.” It’s “has anyone actually looked for AI use that isn’t approved.” Those are very different questions, and only one of them tells you what’s really happening in your organization.

Key takeaways

• Guardrails are the specific technical and procedural limits built into a system before something goes wrong — not a fix applied after.
• Shadow AI is AI use inside your organization with no official approval or visibility — usually not malicious, just unmanaged, and it’s almost certainly happening somewhere in yours right now.
• AI governance is the overall structure of policy, ownership, and oversight meant to catch both of the above.
• The question that actually surfaces shadow AI is “has anyone looked for it,” not “is AI approved here.”
• Real governance needs a named owner, required training, and a working path for people to surface new use cases — not just a policy document nobody’s read.

Want the CPE credit? Take the full lesson on EverydayCPE and earn 0.2 CPE credits: The Risk Words: Guardrails, Shadow AI, and Why "Is AI Approved Here" Is the Wrong Question