What Is Shadow AI and How Should Enterprises Govern It?

Hamid, the security lead at a logistics company, was reading a routine report when one line stopped him. A marketing laptop had sent a large block of text to a public chatbot at 4:50 on a Friday afternoon. He walked over. Selin, the marketing manager, wasn’t hiding anything. “I pasted in the customer list for the spring campaign and asked it to group the people by buying habits. Took two minutes. It was brilliant.”

It was brilliant, and it was also a spreadsheet of names, emails and order histories sitting on a server Hamid’s company had never reviewed. Nobody had done anything malicious. Selin had a deadline, a free tool and no reason to think twice. That’s what shadow AI looks like in most companies: ordinary people, doing ordinary work, with tools nobody approved.

So what is shadow AI?

“Is this a hack?” asked Dev, the finance director, when Hamid told him on Monday. “No,” Hamid said. “It’s shadow AI.” He explained: AI tools staff use at work that the company hasn’t approved and mostly doesn’t know about. The chatbot in a browser tab. The writing assistant. The extension that summarizes pages. The code helper on someone’s personal account. The meeting recorder that joins calls by itself. “It’s shadow IT’s younger cousin,” he said, “except there’s nothing to install, and the first result is useful inside a minute.”

Why it happens

Hamid asked Selin why she hadn’t used the approved tools. Selin winced. “Which ones? I didn’t know there were any. And when I asked IT last year about a similar request, it took six weeks.” That answer explains most shadow AI. People aren’t being reckless. The tools work. The official route is slow, or nobody knows it exists. And the deadline is real. Put someone between the deadline and the process and the deadline tends to win.

What can go wrong

Dev asked the obvious question. “What’s the worst case?” Hamid took a breath. The danger, he said, is that customer data slips out of the company’s hands, particularly on free plans that keep whatever you type or use it to train their models. Terms differ by tool and by plan, so he’d have to check this one. Privacy rules might cover what Selin pasted. And once people trust a tool, its answers land in reports and emails unchecked, so a confident mistake gets passed along. “Anything else?” Dev asked. “Records,” said Hamid. “If nobody knows which tools were used, nobody can say later where a piece of work came from.”

Why a ban doesn’t work

Hamid’s first instinct was to block the chatbot sites. He did, and a week later he spotted the same traffic going through a different tool, and a few staff using their own phones. A ban doesn’t remove the need. It just moves the activity out of sight, and you lose the one thing you had, which was visibility. Companies that handle this well tend to do the opposite. They make the right way easy, and then they make the wrong way less attractive.

What Hamid did instead

Hamid began with a chat, not a rule. For two weeks he wandered between marketing, sales, finance and engineering asking one thing: what are you using, and why? The answers ran longer than he’d guessed, and the reasons made sense. Summaries. First drafts. Spreadsheet help. Quick translations.

In week three he picked a short list of approved tools whose business terms spelled out how data is handled, and sent everyone a link. In week four he wrote a one-page policy and ran a thirty-minute session with real examples, including Selin’s story with the names changed (she volunteered to be in it). He also set up a quick route for new requests, promising an answer within five working days. And from then on he looked at what was actually in use, through the network logs and through the apps staff had connected to company accounts.

What the one-pager said

Dev read the draft over lunch and came back with a pen. His first note was “too short.” Hamid said that was the idea. The page had three parts. Use the approved tools for drafting, summarizing and brainstorming. Never put customer data, passwords, unreleased financials or anything under NDA into a tool that hasn’t been cleared for it. And if you’re not sure, ask in the security channel, where someone answers inside a working day. “Nobody reads twenty pages,” Hamid said. “They file them.” Dev put the pen down.

The conversation with Selin

A month later Hamid met Selin again. She was using an approved tool now. Selin had a new request. She wanted the tool to analyze customer feedback. Fine, Hamid said, as long as the names come out first, and he showed her how. She was pleased. He was relieved. No one had been disciplined. The company simply knew where its data was going now.

Keep the list alive

Three months on, Dev asked whether the approved list needed updating. Hamid had already done it. One tool had changed its data terms in the spring, so it moved off the list. A request he’d turned down in March, a meeting summarizer, had since been given business terms he was comfortable with, so it moved on. He keeps a short log of every decision with the reason beside it. It saves him answering the same question twice, and an auditor will want to see it one day.

Do you have a problem?

Look for the signs. Unfamiliar AI domains in web traffic. Extensions installed on work browsers. Meeting bots appearing on calls. Expense claims for AI subscriptions. Better still, ask, and make it obvious you’re curious rather than accusing. In most companies the honest answer is “more than you think.”v2soft.com/blogs/

If you’d like help setting up a policy or a review, our AI governance assessment services are a good place to start. The guide to AI risk management for enterprise rollouts and the article on enterprise AI governance and compliance give useful background.