You’re Probably Already Using Shadow AI at Work. Here’s What It Can Cost

Somewhere in your company right now, someone is pasting a client email into a chatbot to “make it sound better.” Someone else is uploading a spreadsheet to summarise it. Nobody approved it, nobody logged it, and in most cases nobody meant any harm.

That’s shadow AI: the use of AI tools, accounts or browser extensions at work without the knowledge or approval of your employer’s IT or security team. It’s the AI-era cousin of “shadow IT,” and it’s growing because the tools are free, fast and genuinely useful.

The stakes are no longer theoretical. IBM’s 2025 Cost of a Data Breach research, as summarised by Kiteworks, found shadow AI incidents accounted for 20% of breaches studied, and that those breaches cost about $670,000 more on average than standard incidents. This guide is written for employees, not just security teams: what shadow AI is, what actually goes wrong, what you should never paste into a chatbot, and how to keep the productivity without the risk.

What Is Shadow AI?

Shadow AI covers any AI use outside approved channels. Common examples include:

• Using a personal ChatGPT, Claude, Gemini or Copilot account for work tasks
• Installing AI browser extensions that read the pages you visit
• Uploading documents, code or meeting recordings to an AI summariser
• Using AI note-takers that join calls and send transcripts to a third party
• Connecting AI tools to work email, calendars or cloud storage

The defining feature isn’t the tool; it’s that your organisation can’t see it, assess it or protect the data going into it. A tool that’s approved with a proper business agreement is a very different thing from the same tool used through a personal account.

It’s also not a fringe behaviour. People reach for AI because deadlines are real and the approved options are sometimes slow, limited or non-existent. That’s the pattern worth understanding: shadow AI is usually a symptom of a gap, not a sign of bad employees. For a broader look at which tasks suit automation in the first place, see which tasks are actually safe to automate.

Why Shadow AI Is Risky

Your data leaves the building

When you paste text into a consumer AI tool, that text goes to a third party. Depending on the provider and your account settings, it may be stored, reviewed for safety, or used to improve models. Business and enterprise plans usually come with stronger contractual protections, which is exactly what personal accounts lack. We covered the settings side in the biggest AI privacy risk isn’t a hacker, it’s a setting you never checked.

Breaches get more expensive

Per the Kiteworks summary of IBM’s report, shadow AI breaches averaged $4.63 million versus $3.96 million for standard incidents. The same coverage reported that 65% of shadow AI incidents involved customer personally identifiable information, compared with a 53% global average. In other words, the data that leaks through unapproved AI tends to be the sensitive kind.

Security controls are missing

The report coverage also found that among organisations with AI-related breaches, 97% lacked proper AI access controls, and 63% of breached organisations had no AI governance policy. Shadow AI thrives where there are no clear rules and no approved alternative.

Accuracy and accountability

AI tools can produce confident errors. If an employee pastes in a contract clause, accepts a summary without checking and acts on it, the mistake is theirs, not the model’s. Unapproved tools also leave no audit trail, which matters in regulated industries.

Legal and compliance exposure

Sending personal data, health information, client confidences or source code to an unvetted service can breach privacy law, contract terms or professional duties. New regulation is raising the bar; our plain-English EU AI Act explainer shows why organisations are now expected to know which AI they use.

What Never to Paste Into a Chatbot

If you take only one section from this article, take this one. Unless your employer has approved the specific tool for the purpose, don’t enter:

• Customer or employee personal data: names with contact details, IDs, addresses, account numbers
• Credentials: passwords, API keys, access tokens, private keys, connection strings
• Confidential business information: unreleased financials, strategy, deal terms, pricing, board material
• Source code from proprietary systems
• Legal, HR or medical material: contracts under negotiation, disciplinary cases, health information
• Anything covered by an NDA or client confidentiality agreement
• Full meeting recordings or transcripts with sensitive discussion

A quick test: would I be comfortable if this text appeared in an email to a stranger? If not, don’t paste it. When you need AI help with sensitive material, remove names, figures and identifying details first, or ask whether an approved tool exists.

Real-World Scenarios: Where Shadow AI Sneaks In

Abstract warnings are easy to ignore, so here are four situations that look harmless and aren’t.

The helpful summary. A project manager uploads a 40-page client report to a free AI summariser before a meeting. The summary is good. The report, which included named customers and pricing, now sits on a third party’s servers under consumer terms.

The code assistant. A developer pastes a failing function from a proprietary codebase into a chatbot, including a hard-coded API key they forgot about. The bug is fixed in minutes; the key is now in a chat log and needs to be rotated.

The meeting bot. Someone adds an AI note-taker to a call to save time. It records and transcribes everyone, including a client who never agreed to be recorded, and emails the transcript to a personal address.

The browser extension. An “AI writing helper” extension is installed for convenience. It can read every page the employee opens, including internal dashboards and webmail.

In each case the person’s intent was productivity. The risk came from invisible data flow, which is why the fix starts with visibility and approved tools rather than blame.

A Simple Self-Check Before You Use an AI Tool

Before you paste or upload anything, run through four questions. Is this tool approved for work use? Am I signed in with my work account rather than a personal one? Does the content include personal data, credentials or confidential material? And could I do this task just as well with the sensitive parts removed? If any answer is uncertain, pause and ask. A thirty-second message to IT is cheaper than an incident report.

How to Use AI at Work Without Going Rogue

You don’t need to stop using AI. You need to use it in the open. Here’s a sensible sequence.

1. Find your policy. Search your intranet or ask IT whether there’s an AI acceptable-use policy and a list of approved tools.
2. Ask before you assume. If nothing exists, send a short message: which tools are allowed, and for what kinds of data?
3. Use work accounts, not personal ones, for approved tools, so your organisation’s data protections apply.
4. Check the settings. Where you have the option, turn off training on your data and disable chat history you don’t need.
5. Sanitise your inputs. Replace real names and numbers with placeholders before pasting.
6. Verify the output. Treat AI text as a draft from a fast but fallible assistant, and check facts, figures and citations.
7. Disclose when it matters. If your team or client expects to know AI was used, say so.

If you want to understand where AI judgement stops and human accountability begins, the real reason some decisions can never go to AI is a useful companion read.

For Managers and Leaders: Fix the Gap, Don’t Just Ban

Blanket bans rarely work. If people have deadlines and no approved option, usage moves onto personal phones and home laptops, where you can see even less.

Better approaches start with a clear, short policy written in plain language: what’s allowed, what’s not, and who to ask. Provide a sanctioned tool with proper contracts so employees have a safe default. Train people with real examples rather than abstract warnings. And be transparent about any monitoring; it helps to know what your employer can legally monitor and what they can’t, from both sides of the table.

Given the IBM findings on missing access controls and governance policies, the most valuable first steps are unglamorous: inventory which AI tools are in use, set access rules, and decide who owns AI risk.

Frequently Asked Questions

Conclusion

Shadow AI isn’t a story about reckless employees. It’s what happens when powerful tools arrive faster than policies do. The IBM-based figures are a useful warning: where governance and access controls are missing, the unapproved use that feels harmless can turn into a costly breach.

The fix is mostly habit. Know your policy, use approved accounts, keep sensitive data out of chatbots, and check what the AI gives you. Do that, and you keep the speed AI offers without becoming the weak link in your organisation’s security.