Shadow AI: How to Find and Govern the AI Your Company Can't See
November 10, 2026
How many AI tools, models and agents are running in your organization right now?
If the honest answer is "we're not sure," you are not alone. Employees paste documents into public chatbots. Business teams connect AI assistants to SaaS apps. Developers spin up agents that call internal APIs. Vendors quietly add generative AI features to software you already pay for.
Each decision makes sense locally. Together, they create shadow AI: AI usage that IT, security and risk teams cannot see, cannot govern and cannot defend when something goes wrong.
This article explains why shadow AI has become a board-level risk, where it hides, what it costs, and how to build a practical discovery and governance model that keeps innovation moving.
Shadow AI is any use of artificial intelligence in the organization that happens outside approved governance: no security review, no data classification, no owner and no record in an inventory.
It shares DNA with shadow IT, but three differences make it more dangerous:
That last point matters. Shadow AI is not only a behavior problem. It is a visibility problem across people, applications, vendors and infrastructure.
The data from 2026 shows that shadow AI has moved from anecdote to measurable exposure.
The Lenovo Work Reborn Research Series 2026, a global survey of 6,000 enterprise employees, found that between one-fifth and one-third of workers use AI outside the influence and governance of the IT function.
A 2026 survey by Wakefield Research, commissioned by PagerDuty among 1,250 office professionals at companies with at least USD 500 million in annual revenue, reported that 66% had used AI tools at work despite believing it was not permitted by company policy. In the same study, 39% said they would rather use AI without telling anyone, and 34% admitted entering customer data into public AI tools.
These numbers come from different methods and audiences, so they should not be added together. They point in the same direction: a significant share of AI usage happens where governance cannot see it.
IBM's Cost of a Data Breach Report 2026, based on 602 organizations, puts the global average cost of a breach at USD 4.99 million, a record. According to analyses of the report, 43% of security incidents involved unapproved AI tools, up from 20% the year before, and breaches involving shadow AI averaged USD 5.39 million.
The report also found that AI-related breaches reached 21% of breaches, and that 92% of organizations with an AI-related incident lacked AI access controls.
ServiceNow's 2026 Enterprise AI Maturity Index, a ThoughtLab survey of 4,500 executives in 19 countries, shows a sharp divide. Among the 21% of organizations it classifies as Pacesetters, 61% have AI testing, auditing and risk assessment processes in place. Among everyone else, the figure is 10%.
The gap between how fast AI spreads and how fast governance catches up is exactly where shadow AI grows.
To find shadow AI, you need to look in more places than a typical shadow IT review covers.
Browser-based chatbots and writing assistants used with corporate data. This is the most visible category and the one most policies focus on.
CRM, collaboration, HR, design and productivity tools increasingly include AI capabilities that can be switched on by an administrator or even an end user. They inherit the original app's approval without a new risk review.
Teams build agents with low-code tools, scripts or AI frameworks that connect to internal systems. They may run under personal credentials or service accounts with broad access.
Data science and development teams consume model APIs and deploy models across cloud environments. Without discovery, nobody knows which models process which data.
Suppliers process your data with their own AI. This belongs in third-party risk management, and it rarely gets the scrutiny it needs. If your vendor risk process is already strained, our article on vendor risk compliance shows how to bring it under control.
Individual subscriptions paid by credit card and expensed. They are a useful signal: follow the money and you will often find the usage. The same discipline used in software license audits applies to AI subscriptions.
Shadow AI creates five categories of risk. Each one has a different owner, which is part of why the problem falls between the cracks.
Data exposure. Confidential documents, customer records, source code and financial information leave controlled environments through prompts and integrations.
Regulatory and contractual breach. Data protection laws, sector regulation and customer contracts often restrict where data can be processed and by whom. Unknown AI usage makes compliance impossible to prove.
Uncontrolled actions. Agents with excessive permissions can modify records, send communications or trigger transactions without the controls you apply to people.
Decision quality. Outputs from unvalidated models feed into reports, customer responses and operational decisions with no testing, monitoring or accountability.
Cost sprawl. Duplicate subscriptions, unmanaged API consumption and overlapping tools inflate spend without a clear business case.
None of these risks require bad intent. Most shadow AI comes from people trying to do their jobs faster. That is why banning AI rarely works: it pushes usage further out of sight.
Most organizations started with an acceptable use policy for AI. It is a necessary step, and it is far from sufficient.
A policy tells people what they should do. It does not tell you what they are actually doing. The PagerDuty survey illustrates the gap: 86% of respondents believed their company had formal AI policies, yet two-thirds had used AI in ways they believed were not permitted.
Effective governance needs three things a policy cannot provide on its own:
You do not need to solve everything at once. This five-step framework moves from visibility to control without stopping the business.
Start by finding what exists. Combine several sources:
The tone matters. Position discovery as enablement, not as a hunt. People share more when they believe the goal is to give them safe tools.
Record every AI asset in one place, with the same rigor you apply to other configuration items. A useful AI inventory captures:
If you already maintain a CMDB, extend it rather than starting a separate spreadsheet. AI agents depend on accurate data about services and dependencies, as we explain in CMDB & AI Agents in ServiceNow.
Not every AI use carries the same risk. Create simple tiers:
Each tier gets proportional controls. This keeps governance from becoming a bottleneck for low-risk experimentation.
Shadow AI shrinks fastest when people have approved tools that work well. For each common use case you discover, decide whether to approve the existing tool with controls, replace it with an enterprise option, or retire it.
Governance is a cycle, not a project. Build routines for:
Our article on AI governance in ITSM covers how to structure policies and roles, and best practices for deploying AI in ITSM details the change management side.
Abstract risk categories are easier to act on when you can picture them in your own organization. These are typical patterns we see across business functions.
Reps paste account notes and pricing discussions into public assistants to draft proposals. Marketing teams connect AI writing tools to the CRM to personalize campaigns. The data involved often includes customer names, contract values and pipeline details.
Recruiters summarize CVs with AI tools, and HR partners draft performance reviews or policy responses with assistants that were never assessed. This is one of the most sensitive areas, because the data concerns individuals and decisions that affect their careers.
Analysts upload spreadsheets to AI tools to build forecasts, explain variances or prepare board materials. Unreleased financial results and forecasts are exactly the kind of information that should never leave controlled environments.
Developers use coding assistants and agents connected to repositories, test environments and sometimes production APIs. Source code, secrets and architecture details can be exposed, and agents can act with permissions far beyond what a single task requires.
Teams experiment with AI to draft responses or summarize cases, sometimes outside the approved service platform. Customer data and complaint details travel with every prompt.
Engineers build automations and agents to triage alerts or run scripts. They are often the most technically capable users, which makes their shadow AI the most powerful and the hardest to see.
Mapping these patterns helps you prioritize discovery where the most sensitive data and the most powerful actions meet.
Governance succeeds or fails on trust. If employees believe that admitting AI use will get them in trouble, usage simply goes deeper underground. The PagerDuty survey found that among those who concealed their AI use, 33% did so to avoid scrutiny from managers or leadership, and 29% were unsure whether their use was even permitted.
A few principles help:
Leaders sometimes see AI governance as a cost center that slows innovation. The evidence suggests the opposite.
It reduces breach exposure. Shadow AI incidents carry a higher average cost, and a large share of AI-related incidents happen where access controls are missing.
It cuts wasted spend. Discovery usually reveals duplicate subscriptions and overlapping tools that can be consolidated.
It speeds up adoption. When teams know which tools are approved and how to request new ones, they adopt AI with confidence instead of hesitation.
It strengthens audit and regulatory readiness. An inventory with owners, risk tiers and approvals is the evidence regulators and auditors increasingly expect.
The ServiceNow AI Maturity Index frames it clearly: its Pacesetters treat AI governance as a competitive advantage rather than a compliance burden, and they move faster because the foundations are in place.
Agents raise the stakes because they act on their own. Governing them requires questions that classic software governance never had to answer:
Some organizations are already treating agents like members of the workforce. In a ServiceNow announcement from October 2026, the CISO of LTM described deploying 1,500 AI agents, each assigned an employee ID and a manager. Whether or not you go that far, the principle holds: every agent needs an owner, a scope and a record.
As AI spreads, spreadsheets and point tools stop scaling. When evaluating a platform to govern AI, look for these capabilities:
This is the space ServiceNow addresses with AI Control Tower. At Knowledge 2026, ServiceNow described its expanded AI Control Tower as an end-to-end solution to discover, observe, govern and secure enterprise AI, with visibility across AI systems, agents and identities from multiple model providers and clouds. In October 2026, the company announced additional capabilities, including shadow AI detection to find unapproved agents on the network and AI Passport, a signed credential that defines what each agent is authorized to do at runtime.
The value of anchoring AI governance on the same platform that runs IT, risk and asset workflows is practical: AI assets become part of the processes you already use for change, access, incidents, vendor risk and audits, instead of living in yet another silo.
If you need to show progress quickly, this plan creates visibility and early control.
Weeks 1 and 2. Name an executive sponsor and a cross-functional group from IT, security, risk, legal and the business. Agree on risk tiers.
Weeks 3 and 4. Run discovery across network, SaaS, cloud and expenses. Launch the team survey.
Weeks 5 and 6. Build the first inventory. Classify the top findings by risk. Identify the five most common use cases.
Weeks 7 and 8. Approve or replace tools for those use cases. Apply controls to high-risk findings. Report the baseline to leadership.
At the end of 60 days you will not have perfect governance. You will have something far more valuable than a policy: a factual picture of your AI footprint and a repeatable process to manage it.
No. It is also a compliance, cost and decision-quality problem. That is why governance should involve IT, security, risk, legal and business leaders together.
Blocking without alternatives tends to push usage out of sight. A better approach is to provide sanctioned tools for common needs and apply controls proportional to risk.
Continuously, if possible. New tools, features and agents appear every week. A one-time scan becomes outdated almost immediately.
Ownership is usually shared, with a designated executive accountable for the program and clear owners for each AI asset.
Shadow AI is a symptom of demand. Your people want to work faster, and AI helps them do it. The goal is to make that usage visible, safe and measurable.
Organizations that get this right gain a double advantage: lower risk and faster, more confident AI adoption, because teams know which tools are approved and leaders know where value is being created.
GB Advisors helps enterprises build AI governance on ServiceNow, from discovery and inventory to risk workflows and agent oversight with AI Control Tower. We can start with an assessment of your current AI footprint and governance maturity.
Want to know how much AI is running in your organization today? Talk to our team and schedule a conversation with a ServiceNow AI governance specialist.
