BLOG

What is shadow AI and how is it different from shadow IT?

Shadow IT, the use of unauthorized software and cloud services, has been a security challenge for over a decade. Shadow AI is its faster, harder-to-detect successor, and it carries risks that shadow IT never did.

Traditional shadow IT involves an employee storing files in a personal Dropbox account or using an unapproved project management tool. The data is in one place. The access is relatively static. Detection, while imperfect, is achievable through network monitoring and endpoint controls.

Shadow AI is different in three critical ways. First, it processes data rather than just storing it, customer records, legal documents, financial models, and source code are fed into external AI systems that train on, log, and retain what they receive. Second, it operates at scale, a single employee using an AI writing tool can expose thousands of data points in a single session. Third, it is behavioural, not technical, it is driven by unmet productivity needs, not malicious intent, which means technical controls alone cannot solve it.

Over 80% of employees use unapproved AI tools, and 665 distinct generative AI applications have been tracked across enterprise environments. The workforce using these tools is not the security team’s adversary, they are employees solving real problems with effective tools that their organization has not provided through approved channels.

How shadow AI enters your organization, and why security teams miss it

The entry points for shadow AI are numerous and largely invisible to traditional security monitoring:

• Web-based AI tools accessed through standard browsers, no software installation required, no endpoint detection signature, no network anomaly unless you are specifically monitoring for AI service endpoints
• Browser extensions with AI capabilities that intercept and process page content, including content from internal systems, client portals, and corporate email, without generating alerts in standard monitoring tools
• Personal free-tier accounts on major AI platforms, Harmonic Security found that 16.9% of sensitive data exposures occurred on personal accounts completely invisible to enterprise IT
• SaaS tools with embedded AI features, productivity suites, CRM platforms, and communication tools that have added AI capabilities employees use without realizing they are sending data to external models
• OAuth integrations between approved tools and AI services, employees grant AI tools access to their email, calendar, and file storage through standard authorization flows that do not trigger security review

The Lenovo Work Reborn Research 2026 report documents a workforce split into two groups: employees equipped with IT-managed tools, training, and oversight, and those operating independently with consumer AI services because their employer has not supplied either approved tools or effective guidance. Shadow AI is largely a supply problem, employees reach for unauthorized tools because sanctioned alternatives are unavailable, slow to approve, or inadequate for the task.

3. Regulatory and compliance violations

In Canada, feeding personal information about individuals into an AI system without appropriate consent or safeguards may violate PIPEDA’s accountability and safeguards principles. Processing data about Quebec residents through unapproved external AI systems may trigger obligations under Quebec Law 25. For organizations in regulated sectors. financial services, healthcare, legal, the compliance exposure from shadow AI is compounding with every session.

4. Intellectual property leakage

Source code, product roadmaps, client lists, and competitive intelligence fed into external AI systems do not stay within your organization’s control. Some AI platforms explicitly state in their terms of service that user inputs may be used for model improvement. Even where data is not retained in training, the transmission represents an uncontrolled disclosure of proprietary information.

5. AI-generated misinformation entering business processes

When employees use unapproved AI tools to generate content, reports, client communications, legal summaries, financial analysis, and that content enters business workflows without review, the organization inherits the risk of AI hallucination, bias, and error. This is a governance and accuracy risk as much as a security one.

Real consequences: what happens when shadow AI causes a breach

Shadow AI incidents do not typically manifest as dramatic breaches. They accumulate quietly, a pattern of data exposure that surfaces during a compliance audit, a regulatory investigation triggered by a customer complaint, or a competitor who knows things they should not know. When they do escalate to a reportable incident, the cost profile is distinctive.

IBM’s 2025 Cost of a Data Breach data shows shadow AI adds an average of $670,000 to breach costs compared to incidents without unauthorized AI involvement. The additional cost comes from legal review of what was exposed and to which AI systems, regulatory analysis of whether notification obligations were triggered, incident response complexity when the exposure vector is not a conventional attack but a pattern of employee behaviour, and the difficulty of establishing exactly what data was processed and retained.

The insider-risk component is equally significant. Mimecast’s State of Human Risk 2026 report found that insider-driven incidents, the category that includes negligent AI use, carry an estimated average cost of $13.1 million per incident, with organizations experiencing roughly six such incidents per month.

How to detect AI tools your organization did not approve

Detection requires visibility across three layers simultaneously:

Organizations supported by managed detection and response services can continuously monitor AI activity, identify anomalous behavior, and respond to emerging threats more quickly.

1. Network traffic analysis, monitoring outbound connections to known AI service endpoints (api.openai.com, api.anthropic.com, gemini.google.com, and the growing catalogue of AI SaaS platforms). This is the most scalable detection method but requires active maintenance of AI endpoint lists as new services emerge.
2. OAuth grant audit, reviewing all active third-party application authorizations across your identity platform (Microsoft Entra, Google Workspace, Okta) to identify AI tools that have been granted access to email, calendar, files, or other corporate data. This surfaces tools that are no longer actively used but still hold live access tokens.
3. Endpoint monitoring, browser extension inventory and application usage monitoring on managed devices can surface AI tools being used through installed extensions rather than web access.
4. Direct employee survey, often the fastest and most revealing method. Frame the question as inventory rather than policy enforcement: ‘Which AI tools do you currently use in your work?’ Employees who are productive with effective tools will tell you about them.

Building an AI acceptable use policy that employees will actually follow

The evidence on shadow AI governance is clear on one point. Many organizations engage cybersecurity consulting services to develop practical AI governance frameworks, acceptable-use policies, and risk management processes that employees can realistically follow.

An effective AI acceptable use policy addresses five things:

1. What data classifications may and may not be entered into AI tools, with specific examples employees recognize from their daily work
2. Which AI tools are approved, for which purposes, and under what conditions
3. How employees request approval for a new AI tool, with a realistic turnaround time that does not incentivize working around the process
4. What the consequences of policy violation are, stated plainly, without ambiguity
5. Where employees go for help, a single point of contact for AI-related questions, not a policy document that ends with ‘contact IT’

The policy should be published alongside approved tool access, not instead of it. Governance without alternatives is friction, and friction is what drives shadow AI in the first place.

Shadow AI risk by department: where to look first

Shadow AI exposure is not evenly distributed across organizations. These are the highest-risk departments based on 2026 incident data and tool adoption patterns:

• Marketing and communications, highest volume of AI tool use, including content generation, image creation, and research tools. High likelihood of client information, campaign strategy, and competitive intelligence exposure.
• Engineering and development, source code fed into AI debugging and code completion tools is one of the most consequential shadow AI exposure categories. Proprietary algorithms, API keys embedded in code, and infrastructure configurations are all at risk.
• Finance, customer payment data, financial models, and acquisition targets are routinely processed through AI tools for analysis and summarization. Regulatory exposure under PIPEDA and Quebec Law 25 is significant.
• Legal and compliance, contract language, litigation strategy, and client confidential information processed through AI summarization tools without legal privilege protection.
• Human resources, employee records, compensation data, and performance information processed through AI tools for drafting, analysis, or communication.

How do I find out which AI tools employees are using without approval?

Start with three sources: an OAuth grant audit across your identity platform to see which AI tools have been authorized to access corporate accounts; outbound network traffic analysis to identify connections to AI service endpoints; and a direct employee survey framed as inventory rather than policy enforcement. The survey is often the fastest method. Expect to find more tools than you anticipate, the average enterprise has approximately 1,200 unofficial AI applications in use.

What is the difference between shadow AI and sanctioned AI?

Sanctioned AI tools have been reviewed and approved by your IT and security teams before deployment. Their data handling, retention, and access permissions are understood, and they are included in your monitoring environment. Shadow AI refers to any AI tool operating without that review, personal ChatGPT accounts, unapproved browser extensions, AI features embedded in SaaS tools, and individually deployed agents. The distinction matters because shadow AI creates data exposure that your security team cannot see and therefore cannot manage.

Does shadow AI use violate data privacy laws in Canada?

It can. PIPEDA requires organizations to implement appropriate safeguards for personal information under their control and to be accountable for how it is handled, including by third parties. Processing personal information about individuals through unsanctioned AI systems, particularly systems whose data retention and training practices are unclear, may constitute a failure of those safeguards. Quebec Law 25 adds additional obligations for any business handling information about Quebec residents, including requirements around automated decision-making and privacy impact assessments for new technology deployments. Organizations should consult legal counsel regarding their specific situation.

What should an AI acceptable use policy include?

At minimum: a clear data classification framework specifying what categories of information (customer data, financial records, source code, legal documents) may and may not be entered into AI tools; a list of approved AI tools and their permitted uses; a process for requesting approval of new tools with a realistic turnaround time; explicit consequences for policy violations; and guidance on where to get help. The policy should be accompanied by access to approved tools, a prohibition without an alternative creates the shadow AI it is trying to prevent.