Monday · July 6, 2026 · Issue #038

I want to start this week by naming something that's almost certainly already happening inside your firm right now — and that most firm leaders have been slow to address because addressing it feels awkward.

Your team is using AI tools you haven't approved. Not because they're reckless. Because they found something that makes their work faster and nobody told them not to.

That's shadow AI. And it's the most common source of AI-related compliance and data exposure in professional services firms today — not a cyberattack, not a bad vendor, not a technical failure. Just people doing their jobs, using the tool that helps them do it faster, without a policy to guide them or a governed alternative to use instead.

This week we're going to deal with it directly. The problem Monday through Wednesday. The fix Thursday and Friday — specifically, how to write an AI policy that actually works, step by step, in plain language your whole team can follow.

Shadow IT has been around for years — employees using Dropbox, personal email, or unauthorized apps for work tasks. IT departments learned to monitor and manage it. Shadow AI is different in a way that matters.

When someone uses unauthorized Dropbox, they store files in an uncontrolled location. That's a risk, but a bounded one. When someone uses unauthorized AI, they're actively sending data to a third-party model — and the data flowing in is often far more sensitive than a file name. A client email. A contract under negotiation. A conversation about a difficult personnel situation. A financial projection that isn't public. It goes in as a prompt. What happens to it after that depends entirely on the tool's data practices — practices most employees have never read and most firms have never reviewed.

There's a second difference. With shadow IT, the risk is mostly about where data is stored. With shadow AI, the risk also includes what the tool does with data — whether it stores prompts, whether it uses them for training, whether it can be queried in ways that surface information from other users' inputs. These aren't hypothetical risks. Samsung famously had an internal code leak when engineers used ChatGPT for work tasks in 2023. The risk profile hasn't shrunk since then — it's grown, as tools have become more capable and more embedded in daily workflows.

And there's the third difference: the prompt itself is intelligence. "Summarize this contract and flag the terms most unfavorable to us" tells the AI — and its operator — not just what the contract says, but what your negotiating concerns are. The question reveals strategy that the document alone doesn't.

Here's what the UpGuard data revealed that I think is the most useful insight for firm leaders: employees who use unapproved tools tend to believe they understand the security requirements well enough to manage the risk themselves. The unauthorized use isn't happening despite awareness — it's happening because people think they know what they're doing.

That changes the management response entirely. If people were circumventing policy out of carelessness, you'd train on risk awareness. But when people believe they're already managing the risk adequately, the problem is that they're using a personal risk tolerance as a substitute for an organizational policy. The fix isn't awareness training. It's giving them a better governed alternative — and making it easier to use than the unauthorized one.

Blanket bans don't work for the same reason. According to a Healthcare Brew survey cited by Unseen Security, providing approved AI alternatives reduced unauthorized AI usage by a significant percentage. (I'm not citing the exact figure because I couldn't verify the primary source directly — but the directional finding is consistent with how shadow IT has historically been managed: access beats restriction.)

The formula that works: a clear policy about what's allowed and what isn't, paired with approved tools that are as easy to use as the unauthorized alternatives. This week we'll build both.