Skip to content

170 practical guides on SAP, cybersecurity and AI, $9.99 each. Shop all books ›

0

Blog›AI

What Is Shadow AI? Risks of Unapproved AI at Work and How to Manage It

Shadow AI is AI used at work without the organization knowing or setting rules for it. Learn how common it is, the data and compliance risks, what happened at Samsung and how to find and manage shadow AI.

Somebody on your team probably used an AI tool this week that your IT department has never heard of. Maybe they pasted a contract into a chatbot to summarize it, connected an AI note-taker to their calendar, or tried a browser extension that rewrites emails. Most of the time it’s harmless and genuinely helpful. Sometimes it means customer data, source code or financial plans end up somewhere the company can’t see or control. That gap has a name: shadow AI.

Shadow AI is the use of artificial intelligence tools, models or features at work without the knowledge, approval or oversight of the organization’s IT, security or governance teams. It includes public chatbots, AI browser extensions, AI features switched on inside approved apps, personal AI accounts used for work and AI agents connected to company systems without review.

In one sentence: Shadow AI is AI that people use for work without the company knowing about it or setting rules for it, which means company data can flow to places nobody is watching.

Key takeaways

  • Shadow AI is widespread. In one 2025 survey, 59% of employees said they use unapproved AI tools at work.
  • IBM’s 2025 Cost of a Data Breach Report found one in five organizations reported a breach involving shadow AI.
  • The main risks are data leakage, compliance breaches, unreliable outputs and unmanaged connections to company systems.
  • Blanket bans tend to push usage further out of sight. Approved alternatives plus clear rules work better.
  • Visibility comes first: you can’t govern AI use you can’t see.

How common is shadow AI?

Very. Surveys vary in method, but they point the same way:

  • Research by Cybernews in August 2025 found that 59% of employees use unapproved AI tools at work. Most of them said their managers either supported it or didn’t care.
  • A survey of 2,000 employees in the UK and US at companies with more than 500 staff, conducted by Sapio Research for BlackFog in November 2025, found that 51% had connected or integrated AI tools with other work systems or apps without IT approval or oversight.
  • IBM’s 2025 Cost of a Data Breach Report found that only 37% of organizations have policies to manage AI or detect shadow AI.

The picture is consistent. People adopt AI because it saves time, and governance hasn’t kept up.

Why does shadow AI happen?

Shadow AI is rarely malicious. The usual reasons:

  • The tools are useful. AI really does speed up writing, summarizing, coding and research.
  • Approved options are missing or worse. If the company offers nothing, or a clunky tool, people find their own.
  • Getting approval is slow. A three-month review for a tool someone wants to try today invites workarounds.
  • AI is built into everything. Many apps people already use have switched on AI features, sometimes by default.
  • The rules are unclear. Many employees honestly don’t know what they’re allowed to put into an AI tool.

Understanding the reasons matters, because the fix has to address them. Telling people “don’t” without offering a better option rarely works.

What are the risks of shadow AI?

Risk What can happen
Data leakage Confidential data, personal data or source code is sent to a service with unknown retention, training or access rules
Compliance breaches Personal or regulated data is processed in ways that break privacy laws, contracts or industry rules
Intellectual property loss Trade secrets or unreleased plans leave the company’s control
Inaccurate outputs Decisions or documents rely on AI answers nobody checked
Unmanaged integrations AI tools and agents get access to email, files or business systems through connections nobody reviewed
Account and vendor risk Personal accounts with weak passwords, or small vendors with weak security, hold company data
Prompt injection AI tools connected to company data can be manipulated by hidden instructions in documents, emails or web pages

IBM’s 2025 report put a number on the cost. Organizations with high levels of shadow AI saw breach costs about $670,000 higher on average than those with low or no shadow AI. Breaches involving shadow AI were also more likely to compromise personal information and intellectual property.

The integration risk is growing fastest. A chatbot that only sees what you paste in is one thing. An AI agent with access to your mailbox, calendar and file storage, connected through something like the Model Context Protocol or an OAuth grant, can read and act on far more. If nobody reviewed that connection, nobody knows what it can do.

What about shadow AI agents?

The next wave of shadow AI goes beyond chat. AI agents can take actions: send emails, update records, move files, run code. Employees can now set up agents with a few clicks in many tools, often by granting broad access to their own accounts.

An unreviewed agent carries every risk of an unreviewed chatbot, plus the ability to act. It runs with the employee’s permissions, so if that employee can approve invoices or edit customer records, so can the agent. It may keep running after the employee changes jobs or leaves. And because agents read content from many sources, they’re exposed to prompt injection from emails and documents they process.

Practical steps for agents:

  • Treat each agent as an identity with an owner, a purpose and an expiry date.
  • Give agents only the access they need, and prefer read-only where possible.
  • Require human approval for actions that move money, delete data or contact outsiders.
  • Log what agents do, and include them in access reviews alongside people.

What happened at Samsung?

One of the most cited early cases came from Samsung in 2023. According to reports at the time, engineers in its semiconductor division used ChatGPT for work shortly after the company allowed it. In separate cases, employees pasted in source code to find bugs and uploaded a meeting transcript to produce notes. That confidential material was sent to an outside service.

Samsung responded by restricting generative AI tools on company devices and networks, and reportedly began building internal AI tools instead. The episode made shadow AI a boardroom topic. It also showed the pattern that still repeats: people weren’t trying to leak anything, they were trying to do their jobs faster.

How do you find shadow AI?

You can’t govern what you can’t see. Common ways to build visibility:

  • Network and web traffic. Secure web gateways, firewalls and cloud access security brokers (CASB) can show which AI services people use.
  • OAuth and app connections. Review which third-party apps have been granted access to email, files and calendars in your identity and collaboration platforms.
  • Browser extensions. Inventory and control extensions on managed browsers, since many AI tools arrive this way.
  • Expense and procurement data. Small AI subscriptions often show up on expense reports before IT hears about them.
  • Feature settings in approved apps. Check which AI features are switched on in the software you already use.
  • Just ask. A short, no-blame survey asking which AI tools people use and why often reveals more than any scanner.

How do you manage shadow AI?

The organizations that handle this well usually combine five things:

  1. Provide approved tools. Offer good AI tools with enterprise terms, single sign-on and data protections, so people don’t need to go elsewhere.
  2. Write a clear, short policy. Say which tools are approved, what data can and can’t go into them, and how to request something new. Use examples people recognize, such as “no customer personal data in public chatbots.”
  3. Make approval fast. A lightweight review for low-risk tools, with deeper checks only where data or access risk is high.
  4. Add technical guardrails. Data loss prevention rules, blocks on high-risk services, controls on OAuth grants and least-privilege access for AI agents. Our guide to identity and access management covers the access side.
  5. Train and keep talking. Teach people how to use AI well and safely, including good prompting habits and checking outputs, and keep the conversation open so people report new tools instead of hiding them.

Blanket bans are tempting, especially after an incident. They tend to drive use onto personal phones and home laptops, where the company has even less visibility. A more useful goal is a “safe yes”: a clear, quick path to using AI in a way the company can stand behind.

What questions should you ask before approving an AI tool?

A simple checklist for reviewing a new AI tool:

Question Why it matters
Does the vendor use our inputs to train its models? Training on your data can expose it to other users in some form
How long is data kept, and where? Retention and location affect privacy and compliance
Is there an enterprise plan with a data processing agreement? Consumer terms rarely protect business data
Does it support single sign-on and admin controls? You need to manage who uses it and remove access when people leave
What does it connect to, and with what permissions? Broad access to mail, files or business systems raises the stakes
Can we see logs of what it does? Audit and incident response need a record
What security certifications does the vendor have? Independent audits give some assurance about how the vendor runs its service

A worked example: the helpful meeting bot

A sales manager signs up for a free AI note-taking service. It joins her video calls, records them, writes summaries and emails them to attendees. To do that, she grants it access to her calendar and email with a few clicks.

Within weeks the bot is joining customer calls, internal pricing discussions and a confidential meeting about a planned reorganization, because it’s attached to her calendar and joins everything automatically. Recordings and transcripts sit with a vendor the company has never reviewed, under consumer terms that allow the vendor to use data to improve its service.

The security team finds it during a routine review of OAuth app grants. They don’t start with discipline. They talk to the manager, who explains that the summaries saved her hours each week. The company already licenses a meeting summary feature in its approved collaboration suite, which nobody had told her about. IT revokes the outside tool’s access, asks the vendor to delete the data under its deletion process, and switches on the approved feature for the sales team. Then it adds note-taking bots to its AI policy, with a short list of approved options.

The fix took a week. The lesson was that people will use AI either way, and the company’s job is to make the safe option the easy one.

Frequently asked questions

What is the difference between shadow AI and shadow IT?

Shadow IT is any technology used without IT approval. Shadow AI is the AI version, with added risks because AI tools take in large amounts of data and can act on it.

Is using ChatGPT at work shadow AI?

It is if your organization hasn’t approved it for that use. Many companies now offer an approved enterprise version, which is not shadow AI when used within the rules.

Should companies ban AI tools?

Blanket bans often push AI use out of sight. Most organizations do better offering approved tools, clear rules on data and fast approval for new requests.

How do you detect shadow AI?

Through web traffic monitoring, reviewing OAuth app grants and browser extensions, checking expense reports and AI features in approved apps, and asking employees directly.

The bottom line

Shadow AI is AI used at work without the organization’s knowledge or rules, and it’s already common. The risks are real, from data leaks to costlier breaches, but most of it comes from people trying to work faster. Build visibility, offer good approved tools, set short and clear rules about data, review AI connections to company systems, and make asking for a new tool easier than going around the process.

Sources: Help Net Security: IBM Cost of a Data Breach Report 2025; TechRepublic: shadow AI and poor governance linked to costlier breaches; Techstrong.ai: Cybernews shadow AI survey; BlackFog shadow AI research; The Register: Samsung generative AI restriction.

Go deeper with these books