Shadow AI Risk: The Compliance Danger of Unapproved AI Tools at Work

According to IBM's 2026 Cost of a Data Breach Report, as summarized by the law firm Baker Donelson, incidents involving shadow AI by employees more than doubled to 43% of those studied, up from 20% the year before. They also cost more: $5.39 million on average, against $4.63 million.

The tool is rarely the problem. The problem is the tool nobody reviewed, and that is the heart of shadow AI risk.

Shadow AI risk is the exposure created when employees use AI tools your company has not reviewed or approved: company data leaves your control, vendor and privacy obligations go unmet, and you cannot show an auditor or a customer which tools were in use. You cannot ban your way out of it. You find it, approve a short list, write the rules down, train people, and keep the evidence.

By the end of this guide, you will be able to run a discovery in 30 minutes, map what you find to the controls it touches, and have an approved-tool list and a policy in place within a month.

Here's what I'll cover:

• What shadow AI is, how it differs from shadow IT, and where it hides
• Why employees reach for unapproved AI tools
• The four risks that matter most, with the IBM 2026 numbers
• What SOC 2, ISO 27001, ISO 42001, GDPR, HIPAA and the EU AI Act each expect
• How to detect shadow AI without a security team
• A 30-day plan, an approved-tool matrix and a five-question intake for new tools
• A worked example, the mistakes to avoid and a closing checklist

What Is Shadow AI? Defining Shadow AI Risk Without the Jargon

The UK's National Cyber Security Centre defines it in one line: shadow AI is "the use of AI technology which isn't captured in an organisation's approved systems and processes" (NCSC blog, 7 September 2026).

In plain terms, it is any AI your organization has not reviewed, approved or cannot see. That covers a personal chatbot account, an AI feature someone switched on inside a tool you already pay for, a browser extension, a meeting notetaker, and a coding agent running on a laptop.

The word "shadow" is about visibility, not intent. Most people using these tools think they are being efficient, not breaking a rule.

Shadow AI vs Shadow IT: Where the Shadow AI Risk Differs

Shadow IT is the older problem: someone signs up for a file-sharing app nobody approved. The shadow AI vs shadow IT distinction matters in practice, because shadow AI looks similar and behaves differently.

The last row is the important one. With shadow IT, the app is the unit of control. With shadow AI, the unit of control is the combination of tool, plan and data, and that is why a simple blocklist falls short.

Shadow AI Examples: Where Unapproved AI Tools Hide

These shadow AI examples account for most of what small teams find. Check each one before you assume you know your exposure.

The case everyone cites is still the clearest. In April 2023, a Samsung engineer uploaded internal source code to ChatGPT, and the company then banned generative AI tools in a memo, as Forbes reported on 2 May 2023. Nobody set out to leak anything. Someone wanted help debugging.

Why Employees Use Unapproved AI Tools, and Why It Keeps Growing

The adoption is not a fringe habit. Microsoft and LinkedIn's 2024 Work Trend Index, a survey of 31,000 knowledge workers, found that 78% of AI users were bringing their own AI tools to work. That is a 2024 figure, and the tools have only become easier to reach since.

The reason is usually a gap between what people need and what the official path offers. One Hacker News commenter, writing in July 2026 about a vendor copilot that could not connect to the tools staff actually use, predicted shadow AI "with ugly consequences for security and data exfiltration" when the sanctioned option falls short.

Four patterns repeat:

• The approved tool is slower or weaker. A free chatbot answers in seconds. An access request takes a week.
• Nobody said no, and nobody said yes. Silence reads as permission.
• The feature just appeared. An existing vendor adds an AI button, and a person clicks it.
• The deadline is today. People optimize for the task in front of them.

The Shadow AI Risk Stack: What Goes Wrong When AI Is Unapproved

Start with what IBM's 2026 report found, again as summarized by Baker Donelson. Shadow AI was involved in 43% of the incidents studied, up from 20% in 2025, and those incidents cost $5.39 million against $4.63 million. Among organizations that suffered an AI-related breach, 92% lacked proper AI access controls and 68% lacked AI governance policies, up from 63% the year before. The sample was 602 organizations, between March 2025 and February 2026.

The numbers describe the outcome. The four risks below describe how you get there.

Shadow AI Risk 1: Data Exposure and Confidentiality Breaches

A paste into a chatbot is a data transfer to a third party. Customer records, source code, credentials and unreleased plans all go through the same text box.

What happens next depends on the plan and the vendor's terms: how long inputs are retained, who can read them, and whether they can be used for training. On a personal account, you did not read those terms and you cannot change them.

Your own customer contracts add a second layer. Most enterprise agreements include confidentiality clauses and data processing terms, and an unreviewed AI tool sits outside both.

Shadow AI Risk 2: Contract, Privacy and Regulatory Breaches

Privacy law cares about who touches personal data, not how convenient the tool was.

• GDPR: a vendor that processes personal data for you is a processor, and Article 28 requires a contract with specific terms. An unapproved AI tool has none. If personal data leaks through one, Article 33 starts a 72-hour clock for notifying the regulator, and you cannot start it if you never knew the tool existed.
• HIPAA: a vendor that handles protected health information needs a business associate agreement before it touches any. A free chatbot does not sign one.

Shadow AI Risk 3: Missing Evidence When an Auditor or Customer Asks

Auditors sample your vendor list and your approvals. An AI tool in daily use that appears on neither is a gap you cannot explain away.

Customers increasingly ask the same thing in security questionnaires: which AI tools touch our data, and where do the prompts go? If the honest answer is "we are not sure," the deal slows down. The same pattern shows up in our guide on whether SOC 2 covers AI tools like ChatGPT and Claude.

Shadow AI Risk 4: Unreviewed Outputs in Customer-Facing Work

Inputs are half the risk. Outputs are the other half. A wrong answer drafted by an unapproved tool and sent to a customer under your name is your error. Code generated by an unreviewed assistant can carry licensing and security problems into production.

Output risk is the reason a good policy includes a human-review rule, not only a data rule.

Shadow AI Compliance: What Each Framework Expects

No major framework mentions shadow AI by name. Shadow AI compliance comes down to controls you already have, and an auditor will find the gaps whether or not the term appears in the standard.

SOC 2 and ISO 27001: Vendor, Acceptable Use and Leakage Controls

SOC 2 has no AI-specific criterion, and the point is not that it requires an AI program. Unreviewed AI vendors fall under criteria you already test: vendor and business-partner risk (CC9.2) and risk assessment (CC3). ISO 27001:2022 is more explicit, because its Annex A already names the relevant controls.

For the ISO side of the stack, our guide on ISO 27001 and AI compliance covers how the 2022 controls extend to AI.

ISO 42001, GDPR, HIPAA and the EU AI Act: The AI-Specific Layer

If you pursue ISO 42001, the AI management system standard, shadow AI cuts directly against the AI policy, supplier and use controls. For the full control list, see our ISO 42001 explainer.

How to Detect Shadow AI Risk in a Company With No Security Team

You do not need a CASB or a network sensor to start. Seven sources, most of them already in your hands, will show you more than you expect.

One caution before you go further. Monitoring has an employee-privacy side. Tell people what you will look at and why, keep the amnesty promise in the survey, and aim the exercise at tools and data, not at individuals.

How to Prevent Shadow AI Risk in 30 Days: A Playbook for Small Teams

This is how to prevent shadow AI without a blanket ban, in the sequence I would run at a company under 50 people. The week labels and effort are my estimates, not benchmarks, and a team with a busy audit calendar should stretch them.

1. Discover, with amnesty (week 1). Run the survey and the two quick checks above. Promise that nobody is in trouble for what they report.
2. Classify the data (weeks 1 to 2). Decide which data classes exist and which AI tools may ever see each one. Our data classification policy guide gives you the tiers.
3. Decide on each tool you found (week 2). Approve, approve with limits, replace, or block. Publish a one-page approved list.
4. Write the rules and the request path (week 3). Put them in your policy, with a named owner and a response-time target for new tool requests.
5. Train and collect acknowledgements (weeks 3 to 4). Short, specific, tied to the approved list.
6. Set the review rhythm (week 4). A quarterly re-scan and an exceptions log.

Why a Ban Alone Fails

The instinct after a scare is to block everything. Samsung did exactly that in 2023, and was reported at the time to be developing its own internal AI tools for the same jobs, which suggests the need did not go away.

A Hacker News commenter put the failure mode plainly in February 2026: trying to ban this outright "will only push your most effective employees" toward places where they can work freely. That matches how I would expect people to behave. If the blocked tool saves two hours a day, it moves to a personal phone, and now you have the same risk with less visibility.

A ban is still the right answer for one thing: the most sensitive data. Keep the hard block for regulated and confidential data, and give people an approved route for everything else.

The Approved-Tool Matrix: Data Class by Tool Status

The matrix answers the question employees actually ask: can I put this in that?

An "approved enterprise tool" here means a company account on a business plan with reviewed terms, not a personal login. The same product can sit in different columns depending on the plan you use.

Shadow AI Policy: What to Put in It

A shadow AI policy is usually a section of your acceptable use policy, not a separate document. The rules text lives in the acceptable use policy guide, which has an AI section covering an approved-tools list, a ban on confidential data in public AI models and human review of AI content, and in the AI governance policy guide, which includes approved-tool and prohibited-use clauses.

Whatever the wording, cover these eight points:

• The approved tool list and who maintains it
• What data may go into which tool (the matrix above)
• How to request a new tool, and how fast you will answer
• Whether personal AI accounts may ever be used for work
• Rules for AI features inside approved tools
• Human review before AI-written content reaches a customer
• How to report a mistake without fear, such as pasting the wrong file
• The consequence for knowingly ignoring the policy

A policy nobody has acknowledged is hard to defend. We offer a pre-built acceptable use template, and the platform logs each employee's acknowledgement with a timestamp and tracks version history. Pair it with the short training described in our security awareness training policy guide.

Vetting a New AI Tool Request: Five Questions

The request path only works if the review is quick. Five questions cover most of it.

Anything that fails the first or third question stays out of the confidential and regulated rows of the matrix. For the longer version of this review, see our vendor management policy guide.

Shadow AI Governance Example: One 25-Person Company's First Month

Shadow AI governance is easier to see in a worked case, so the following example is hypothetical. Larkspur is an invented 25-person B2B SaaS company preparing for SOC 2 Type 1. Every tool and decision below is illustrative, not a benchmark.

After the survey and the two quick checks, Larkspur's lead finds eight tools in use:

Three things stand out in the result.

Most decisions were not blocks. Three of eight were blocks. The rest were approvals, limits or replacements, which is why the plan survived contact with the team.

The riskiest finding was the quietest. The OAuth grant had no invoice and no user complaint. It only appeared because someone read the identity provider's connected-app list.

The evidence is now the deliverable. Larkspur ends the month with an approved list, a matrix, signed acknowledgements and an exceptions log, which is exactly what an auditor will ask for.

Shadow AI Risk Mistakes That Leave the Gap Open

Each of these is easy to make, and each leaves the exposure in place.

• Running a one-time scan. Tools change monthly. A discovery done once in spring is stale by autumn, so schedule the re-scan.
• Blocking without an alternative. If there is no approved way to summarize a call or draft a reply, people will find their own.
• Writing a policy and skipping the acknowledgement. An unsigned policy is a suggestion, and an auditor will treat it as one.
• Ignoring AI features inside approved tools. The vendor is on the list, but the new AI button was never reviewed.
• Treating all data the same. "No AI with company data" is unworkable, so people ignore it. The matrix is easier to follow.
• Forgetting personal accounts. The same product can be fine on a company plan and a problem on a personal login.
• Having no owner. If nobody maintains the approved list and answers requests within days, the process quietly dies.

Shadow AI Risk Checklist

Answer yes or no to each. Every no is a task.

• Have we surveyed staff about AI use, with an amnesty promise?
• Have we checked SSO and OAuth grants for AI apps?
• Have we searched expense and procurement lines for AI subscriptions?
• Have we reviewed AI features inside our approved SaaS tools?
• Do we have a one-page approved AI tool list with an owner?
• Do we have a data-class matrix that says what may go into which tool?
• Is there a named request path for new tools, with a response-time target?
• Does our acceptable use policy cover AI, and has every employee acknowledged it?
• Have we trained staff on the approved tools and the rules?
• Is a quarterly re-scan on the calendar, with an exceptions log?

FAQs

What Is Shadow AI Risk?

Shadow AI risk is the exposure created when employees use AI tools the company has not reviewed or approved. Data can leave your control, contract and privacy obligations can be breached, and you cannot evidence what was in use. The tool itself may be perfectly good. The gap is the missing review.

Is Using ChatGPT at Work a Compliance Violation?

Not automatically. It depends on the plan, the data and your policy. A company account on a business plan with reviewed terms, used for public or internal data, is usually fine. A personal account used to paste customer records or source code is the situation that breaches contracts, privacy duties and your own acceptable use rules.

What Is the Difference Between Shadow AI and Shadow IT?

Shadow IT is any technology used without approval. Shadow AI is the subset where the tool ingests your data, may retain or learn from it, and produces outputs that enter real work. That is why a blocklist handles shadow IT better than it handles shadow AI.

How Do You Detect Shadow AI in Your Company?

Start with an anonymous survey, your identity provider's connected-app list, and an export of expenses. Add a review of AI features inside the SaaS tools you already approved. Network or CASB logs help if you already have them, but you do not need them to begin.

Should You Ban ChatGPT at Work to Control Shadow AI Risk?

A blanket ban tends to push use onto personal devices, where you have less visibility. Ban the most sensitive data classes, and give people an approved tool and a fast request path for everything else. That removes the reason to work around you.

How Do You Stop Employees Using Unapproved AI Tools?

You reduce the reasons to use them. Publish an approved list, answer requests within days, train people on what data goes where, and keep a quarterly re-scan. Enforcement matters, but the quickest improvement usually comes from making the approved path faster than the unapproved one.

Does Shadow AI Risk Affect SOC 2 or ISO 27001?

Yes, indirectly. SOC 2 has no AI-specific criterion, but unreviewed AI vendors fall under vendor and business-partner risk (CC9.2) and risk assessment (CC3). ISO 27001:2022 has controls for inventory, acceptable use, suppliers, awareness and data leakage that unapproved AI tools commonly breach.

What Should a Shadow AI Policy Include?

An approved tool list with an owner, rules for which data goes into which tool, a request path for new tools, a position on personal accounts, rules for AI features in approved tools, human review of AI content, a no-blame way to report mistakes, and a stated consequence for knowingly ignoring the rules.

Related Reading

• Acceptable use policy: the policy template, including the AI section
• AI governance policy: the program-level policy with approved-tool and prohibited-use clauses
• Does SOC 2 cover AI?: the SOC 2 audit view of AI tools
• AI risk assessment: how to score the risk of one AI system
• AI governance framework: the program the controls above sit inside
• What is ISO 42001?: the AI management system standard
• ISO 27001 and AI compliance: how the 2022 controls extend to AI