Shadow AI: Frank Abagnale Works in Your Company

At 11 pm tonight, someone in your company will paste a client contract into a free AI account and get their work done in four minutes. They are not a criminal. They are Frank Abagnale before the FBI hired him.

In 1920 the United States banned alcohol, and Americans invented the speakeasy. A century later, companies banned ChatGPT, and employees started building their own little AI assistants on personal accounts, usually at 11 pm, with the calm confidence of someone assembling IKEA furniture without the manual.

A 2025 global study by the University of Melbourne and KPMG, covering more than 48,000 people, found that 57% of employees hide their AI use at work, and almost half admit using AI against company policy.

Picture your open space: half the room typing with suspicious confidence, and the moment a manager walks past, browser tabs close faster than a teenager's phone at 1 am.

None of this is new. Dropbox and personal email caused the same panic fifteen years ago, and researchers called it Shadow IT. The habit stayed the same, only the tool has changed.

Is Shadow AI Good for Innovation?

Everyone remembers Catch Me If You Can for the chase. The interesting part is the ending: Frank Abagnale (Di Caprio) spends years forging cheques, Carl Hanratty spends years chasing him, and then the FBI puts him to work on bank fraud. The man who found every hole in the system became the man who closed them.

Your Shadow AI users are in the same position, minus the pilot uniform. The colleague who built a small agent to summarise contracts has done free product discovery: they found a real problem, built a working prototype, and proved people want it.

In innovation terms, that's a prototype with a first user already on board, and it cost you nothing.

The same agent might also be sending client contracts to a free tool that keeps them, which is roughly the corporate equivalent of reading your diary out loud on the bus. Samsung learned this in 2023, when engineers pasted internal source code into ChatGPT. Its memo restricted generative AI but called the restriction temporary, until the company could build a secure way to use it.

So Shadow AI is a very good signal and a very bad system. It shows you exactly where people need AI, and it handles your data like a teenager handles the family car keys.

What the Academic Research Tells Us

Shadow IT has a surprisingly rich academic literature, built mostly by information systems researchers in Germany and Switzerland over the last fifteen years. Three findings matter directly for AI strategy:

1. Why Good Employees Cross the Line (Steffi Haag & Andreas Eckhardt)

The first explains why good employees cross the line. Steffi Haag and Andreas Eckhardt, two of the field's leading researchers, showed that people make unapproved tools feel acceptable with three excuses: necessity, harmlessness, and injustice.

In plain words: "I had no choice, nobody gets hurt, and the official tool is ridiculous anyway." Users also find these excuses more convincing when they see colleagues doing the same, which is the peer-reviewed version of "but Mum, everyone else is doing it."

Mario Silic and his colleagues added my favourite excuse, which you know from the gym: I went on Monday, so Friday's pizza doesn't count. At work it sounds like: "I closed the quarter, so I've earned a little ChatGPT." In their study, it was the excuse that predicted what people actually did, not just what they said, and they found shame plays a role in holding people back. Every parent already knew that one: the disappointed face does more than the grounding.

2. What Companies Do Once They Discover a Shadow System

The second finding is about what companies do once they discover a shadow system. Researchers identified four possible outcomes: phase it out, replace it, bring it under IT, or let it continue.

Think of finding out your teenager built a treehouse without asking. You can tear it down, build a proper one, have it inspected and officially adopt it, or pretend you never saw it. In a study of three companies, the company chose to rebuild the discovered tool or hand the work over to IT in 62% of cases, so tearing it down was the exception.

The literature even has a name for the healthy destination, "business-managed IT": tools built by the business, but in the open and aligned with IT.

3. Bridges the Gap to AI (Mario Silic, 2025)

The third finding bridges the gap to AI. In 2025 Silic studied Shadow AI directly, through a survey of 140 professionals and interviews with ten executives. He argues it grows from the same roots as Shadow IT but adds new risks, from hallucination (the AI kind, not the Friday afternoon kind) to what he calls governance drift, where nobody knows anymore which decisions were shaped by which tool.

The Paved Road, in Six Steps

Prohibition created speakeasies. If you want a different result, build a paved road: a legal route that is easier than the illegal one.

Step 1: Run an Amnesty

Run it like a library amnesty week for overdue books. Ask every team which AI tools and homemade agents they use, for what, and with which data, and promise nobody gets punished for the answer.

The UK's National Cyber Security Centre recommends exactly this for Shadow IT, for a simple reason: once staff see a colleague blamed, everyone else suddenly develops amnesia about their own workarounds, and you end up governing a map full of blank spaces.

Step 2: Sort by Data, Not by Tool

A traffic-light rule works because anyone can remember it:

The risk is almost never the tool itself. It's what people put into it, and on which account. The same chatbot is a different object depending on whether someone signed up with a personal email or your company pays for it. A generic ban treats a LinkedIn draft and a client database as the same crime.

Step 3: Make the Legal Road Better Than the Shadow One

Give everyone an enterprise AI assistant, and give builders a sandbox where they can create small agents with approved models and safe data. This takes away the necessity excuse.

The tier matters here more than the brand. Free consumer plans generally give you no control over your data and may use it to improve the models, while the business, team and enterprise plans are where you get the contractual commitment not to train on your content, administrator control, single sign-on, audit logs and a data processing agreement your legal team can actually accept.

A few hundred euros a month per team is a small price against one client contract sitting on someone's personal account. Read what you sign rather than trusting the marketing page, because the guarantees differ between vendors and between tiers. The same UK agency advises avoiding unnecessary lockdowns and anticipating what people need, because that may stop the shadow before it starts. A secure tool nobody wants to use just reopens the speakeasy.

Step 4: Filter Data, Not Websites

Blocking chatgpt.com at the firewall simply moves usage to personal phones, where you see nothing at all. Data loss prevention (DLP) tools can instead catch client files, card numbers or code being pasted into external AI services, and warn or block at that moment. It's the difference between banning kitchens and installing a smoke detector.

Step 5: Open a Fast Lane for Homemade Agents

When someone builds something useful, they register it on a single page: what it does, what data it touches, and who owns it.

A small group then decides within days between the four outcomes from the research: phase it out, replace it, bring it under IT, or let it continue as business-managed IT. Speed matters as much as the decision. If registration takes six months, you've recreated the injustice excuse with a form attached.

Step 6: Train People and Reward the Builders

In Europe, the AI Act asks every company using AI to take measures supporting its staff's AI literacy. The rules were softened in July 2026, but the duty is still there, and training is the only real answer to the harmlessness excuse.

Then turn the gym excuse around and publicly credit the people whose agents get adopted. Since shame plays a role in hiding, make declaring your homemade agent something people are proud of.

Governance should be small and fast: a handful of people from IT, security, legal and the business, meeting often, with a registry of agents reviewed every quarter. If your AI governance committee meets less often than your employees build new agents, the shadow wins by default.

Prohibition lasted thirteen years and ended with the bars reopening, this time with licences, taxes and someone at the door checking IDs. Your employees have already opened the bar, and they are not going to close it because you sent a memo. Your job now is to issue the licences.

Abagnale ended up with a badge, after all.