Shadow AI Is the New Shadow IT

A pattern you already know

• ThenYour file server is slow and the VPN keeps dropping, so files end up on personal Dropbox accounts.
• LaterAdding one field to the official CRM takes months, so sales builds its pipeline in Airtable.
• NextSomeone adds up the software paid on corporate cards and finds dozens of subscriptions IT never approved.
• NowThe request for an approved AI tool is still "under review", so people paste emails, code and figures into whatever chatbot is open in the next tab.

None of these scenes comes from one company; most IT teams have lived at least two of them. The tools change every few years. The reason people reach for them stays the same: the approved route is slower than the job allows.

Shadow IT, the first round

Shadow IT covers every tool, service or account used for work without IT knowing about it. You'll recognize most of these:

• Dropbox, Google Drive or WeTransfer for files "too big for email"
• WhatsApp groups and private Slack workspaces for "quick coordination"
• Trello, Notion or Airtable, set up by a team that got tired of waiting
• Marketing analytics, sales prospecting and HR survey tools on corporate cards, none of them in the IT inventory

Dropbox works and Notion is good software. The trouble is the questions you can't answer about them: which data lives there, who can open it, what happens when the person who set it up leaves, and whether any of it breaches a client contract or GDPR.

Shadow AI moves faster

Shadow AI follows the same path, with one difference that matters a lot. The exposure happens the moment someone presses Enter.

A file sitting in an unmanaged Dropbox folder is a risk you can still clean up. A prompt sent to a free AI tool has already been processed, possibly logged, and on some consumer plans used to train the next model. There is nothing to recall.

In practice, it looks like this:

• Personal ChatGPT accounts used to rewrite client emails, summarize contracts or debug production code
• Browser extensions that read every page you open so they can "help you write"
• Free tools whose terms let them reuse what you type
• AI features switched on inside Notion, Slack or the CRM, with nobody checking where the data goes
• Developers running Copilot, Cursor or similar on personal licenses, without any security review

The reasons people give

Nobody starts the day planning to leak client data. They are trying to finish their work, and the official route keeps losing to the deadline. Every IT team hears some version of these:

From where the employee sits, each of these is a rational choice. They see the time saved every single day. The risk shows up two years later in a breach report, usually with someone else's name on it.

The bill comes later

Four costs, roughly in order of how much they hurt:

Banning backfires

The reflex is to block ChatGPT at the firewall and send a stern email. It feels like control. The usual sequence after that:

1. IT blocks AI tools on the corporate network.
2. Staff switch to their phones and home laptops.
3. You lose the little visibility you had.
4. IT becomes the department people quietly work around.

The need hasn't gone anywhere. It has moved to devices you don't manage.

Six things that work

You get further by giving people a safe way to do what they're already doing.

1. Ship an approved option quickly

   A business AI plan with a data processing agreement, one approved file-sharing service, one sanctioned project tool. If the approved tool is also the easiest one, most people will use it.
2. Answer tool requests within 48 hours

   A quick "no, use this instead" beats six months of silence.
3. Measure before you decide

   Read twelve months of expense reports, check DNS logs for AI domains, or use a SaaS discovery tool. Find out what is actually in use first.
4. Explain the real risk

   Fifteen minutes on what happens to a prompt pasted into a free tool does more than any blocked URL. It also covers part of your AI Act Article 4 obligation.
5. Write a one-page AI policy

   Which tools are approved, which data never goes into any AI tool, how to ask for something new. Keep it to one page so people actually read it.
6. Run an amnesty

   Give people a month to declare what they use, with no consequences. You want to hear about the Airtable full of client data from its owner, before a breach notification tells you.

Change how IT says yes

Shadow AI looks like a security problem, so it tends to land on the security team's desk. The fix usually sits with IT service delivery: how fast you approve, how clearly you explain the rules, and whether the approved tools are good enough to use.

Shadow AI shrinks once the approved route is faster than the workaround.

Start this week

Two things. Ask finance for every software and AI subscription expensed in the last twelve months. Then publish a one-page AI policy that names at least one approved tool. I'd bet the expense list teaches you more than any staff survey would.