What Is Shadow AI? A Practical AI Governance Guide for Executives

Maikel Pereira
November 3, 2026

Most executives picture AI adoption as something they will roll out: a strategy, a pilot, a launch. In most companies, it has already started without them, in tools nobody approved and on accounts nobody manages. This is shadow AI, and leadership is part of it. In a 2025 survey of more than 1,000 U.S. employees, 59% said they use AI tools their company hasn’t sanctioned, and among executives and senior managers the figure was 93%.¹ Sometimes what goes into those tools is sensitive: in 2023, Samsung discovered that an engineer had uploaded internal source code to ChatGPT.²

Adoption is not the problem. McKinsey’s 2025 global survey found that 88% of organizations now use AI regularly in at least one business function.³ Ownership is. In McKinsey’s March 2025 survey of 1,491 participants across 101 countries, only 28% of respondents said their CEO is responsible for overseeing AI governance, and just 17% said their board is.⁴ The use of AI has grown faster than anyone’s authority over it.

This article looks at shadow AI and AI governance from the executive’s chair. It covers what shadow AI looks like inside a company, why banning it fails, what it costs when it goes wrong, and five decisions that only leadership can make to bring AI use into the open without slowing teams down. It does not cover model architecture; we cover agent infrastructure and production evaluation in separate articles on this blog.

What shadow AI looks like from the executive’s chair

Shadow AI is any AI tool or feature used for work without the organization’s knowledge or approval: a free chatbot account, a browser extension that summarizes email, an AI feature switched on inside a SaaS product, a coding assistant on a personal license.

The common mistake is to assume it is rare, or to treat it as an IT problem that security will handle. The numbers say otherwise. UpGuard’s 2025 research found that 81% of employees and 88% of security leaders use unapproved AI tools.⁵ BlackFog’s early-2026 survey of 2,000 workers in the UK and US found that 49% use AI tools their employer hasn’t sanctioned, that 60% would take security risks to meet deadlines, and that 69% of presidents and C-suite respondents accept this, prioritizing speed over security.⁶ When the people responsible for security and the people at the top of the company are both using unapproved tools, a policy written by IT alone will not change behavior.

The BlackFog finding also explains the motive: most shadow AI comes from people trying to get work done faster. That makes it a demand signal. It shows where employees see value in AI, which tasks they want help with, and where the approved tools fall short. An executive who reads it that way gets a map of the company’s real AI needs.

Why banning shadow AI fails: the Samsung lesson

The first reaction to shadow AI is often a ban. Samsung Electronics is one widely reported example. After the source code upload, the company issued a memo banning generative AI tools for employees, citing concern that data shared with chatbots is stored on external servers with no easy way to access or delete it. The same report noted that Amazon had warned staff not to share confidential information with ChatGPT, and that JPMorgan Chase and several other large U.S. banks had restricted its use.⁷

A ban can be a reasonable emergency measure. As a long-term policy, it tends to push usage out of sight. UpGuard found that 45% of workers find workarounds to access blocked applications.⁸ In the Cybernews survey, 52% of employees said their employer provides approved AI tools, but only 33% said those tools fully meet their needs.⁹ When the sanctioned option is worse than the free one, people find another way in, and the company loses the little visibility it had.

The fix is to replace the ban with a choice employees prefer: approved tools, clear rules on what data can go where, and a fast way to ask for something new. The five decisions later in this article describe how.

What shadow AI costs when it goes wrong

Shadow AI is often discussed as a vague risk. IBM’s 2025 Cost of a Data Breach Report put a number on it. One in five organizations studied reported a breach involving shadow AI, and breaches at organizations with high levels of shadow AI cost an average of USD 670,000 more than those with little or none, against a global average breach cost of USD 4.44 million. These incidents took about a week longer than average to detect and contain, and were more likely to expose personally identifiable information (65%) and intellectual property (40%).¹⁰

The same report shows how the governance gap turns shadow AI into losses. 63% of breached organizations had no AI governance policy or were still developing one. Among organizations with policies, fewer than half had approval processes for AI deployments, and only 34% performed regular audits for unsanctioned AI. 97% of organizations that reported an AI-related security incident lacked proper AI access controls.¹¹

Governance also shows up on the upside. McKinsey’s research found that CEO oversight of AI governance is one of the factors most correlated with higher reported bottom-line impact from generative AI.¹² Correlation is not proof, but it suggests that governing AI and getting value from it go together.

Decision one: name an owner

The common mistake is to treat AI governance as a compliance document: legal drafts a policy, IT blocks a few websites, and everyone moves on. That gives the appearance of control without the substance. According to NACD data cited by McKinsey in December 2025, fewer than 25% of companies have board-approved, structured AI policies and only about 15% of boards receive any AI-related metrics. McKinsey also notes that only 39% of Fortune 100 companies disclosed any form of board oversight of AI.¹³

The fix is to name a single accountable executive for AI, with explicit decision rights. That person doesn’t need to be technical. They need the authority to answer three questions for the whole company: which uses are allowed, who approves the exceptions, and who is accountable when an AI-assisted decision goes wrong. This matches NIST’s AI Risk Management Framework, which lists executive leadership taking responsibility for AI risk decisions as a core governance outcome.¹⁴ Shadow AI grows where nobody can answer those questions.

Decision two: take inventory without punishment

You can’t govern what you can’t see, and it is easy to underestimate how much AI is running in a company when most of it never went through procurement. The common mistake is to launch discovery as a crackdown. If people expect discipline, the usage is likely to go further underground.

The better approach is an amnesty-style inventory. Ask every team which AI tools they use, for what tasks, and with what data, and make it clear the goal is to support them. Cross-check the answers with what your security and procurement systems already show: browser extensions, SaaS sign-ups, and expense reports for AI subscriptions. Then repeat the audit on a schedule. IBM’s finding, cited above, that only 34% of organizations with AI policies audit regularly for unsanctioned AI shows how rarely this step happens.

The output is a living use-case inventory: a register of every AI tool and use, its owner, the data it touches, and its business purpose. NIST’s framework treats a resourced AI inventory as a governance requirement in its own right.¹⁵ It is the least glamorous document in AI governance and the one everything else depends on.

Decision three: tier the risk instead of treating all AI the same

Not every AI use deserves the same scrutiny. Asking a chatbot to tighten the wording of an internal email is not the same as using AI to screen job candidates or approve credit. When governance treats them identically, one of two things happens: low-risk productivity gets buried in paperwork and moves back into the shadows, or high-risk decisions get the same light touch as the email.

A three-tier model is usually enough to start:

Tier	Typical uses	Governance required
Green (low risk)	Drafting, summarizing non-sensitive content, brainstorming, coding help on non-production code	Allowed on approved tools; basic training; no approval needed
Yellow (moderate risk)	Customer-facing content, analysis of confidential data, AI features inside business workflows	Named owner, approved tool with enterprise data protections, human review before output leaves the company
Red (high risk)	Decisions about people (hiring, performance, credit, access to services), autonomous agents that take actions, regulated data	Formal review, documented human-in-the-loop, audit trail, periodic re-assessment

The red tier follows the same logic regulators use. Under the EU AI Act, obligations for stand-alone high-risk systems, including AI used in employment and creditworthiness decisions, now apply from December 2, 2027, after the Digital Omnibus on AI deferred the original August 2026 date. Most transparency obligations still apply from August 2026.¹⁶ The deadline moved, but the obligations did not go away, and companies that tier their AI use now will have less to fix later.

For the technical controls behind the red tier (tool permissions, audit trails, human review queues), see our article on AI agents in the enterprise.

Decision four: build a paved road people want to use

The most effective answer to shadow AI is a better official option. A “paved road” is a set of approved, well-integrated AI tools that are easier to use than the shadow alternatives. In practice that means enterprise versions of general-purpose assistants with data-retention protections, AI connected to your own knowledge sources, and a lightweight way to request a new tool. If approval takes three months, the free tool wins in three minutes.

We saw the same principle outside AI. When we replatformed more than 3,000 franchise websites on dotCMS for a large U.S. retail company, the platform’s default editing interface wasn’t intuitive for the local partners who needed to make small, frequent updates. Rather than forcing partners to adapt to it, we built a simpler editing layer that gave them safe freedom to update their own sites while governance and brand control stayed central.¹⁷ AI governance works the same way: central control over the data and decisions that matter, and an easy, approved path for everything else.

Decision five: report value and risk on the same page

The last common mistake is splitting the conversation. The innovation team reports productivity wins, the risk team reports incidents, and nobody at the top sees the trade-off. That’s how companies end up over-investing in AI that doesn’t pay off, or shutting down AI that does after a single incident.

A quarterly AI scorecard for the leadership team and the board fixes this. It should cover:

Adoption: how many approved tools and use cases are active, and the estimated share of AI use happening on the paved road versus in the shadows.
Value: business outcomes per major use case (time saved, cycle time, revenue influenced, cost avoided), measured against a baseline. Our article on evaluating AI in production goes deeper on outcome-based measurement.
Risk: incidents, near misses, shadow AI findings from the latest audit, and red-tier use cases pending review.
Cost: AI spend by use case, so value can be judged against unit economics. Our piece on the hidden cost of running AI at scale covers this in detail.

Few companies are there yet. McKinsey’s “State of AI trust in 2026” survey of roughly 500 organizations found that only about 30% reach a maturity level of three or higher in strategy, governance and agentic AI controls, even as average responsible-AI maturity rose to 2.3 from 2.0.¹⁸ A scorecard reviewed every quarter is how an executive turns a policy document into a management routine.

What this looks like in practice: JPMorganChase, from restriction to a paved road

JPMorganChase is a useful example because it went through both stages. In February 2023, the bank temporarily restricted employees’ use of ChatGPT across its global staff. According to a person familiar with the matter, the decision was not a response to a specific incident but part of its normal limits on third-party software for compliance reasons.¹⁹ The bank did not leave the restriction as its long-term answer. It built an approved alternative.

That alternative is LLM Suite, an internal portal that gives employees access to popular large language models. By May 2025 it had about 200,000 users, and American Banker described JPMorganChase as the first big bank to roll out generative AI to almost all of its employees.²⁰

The design reflects several of the decisions in this article. There is a visible executive owner: the effort is led by chief analytics officer Derek Waldron and his team. Data protection is built in: the bank does not let its data flow into public-facing models, and third parties may not train their models on its proprietary data, including the content of employee prompts. New models are tested for safety, security and fit with specific use cases before they are added, and each data source is checked for ownership and authority before it is connected. The tools were also designed to be self-service and “consumer grade,” so the approved option competes with free tools on convenience. Chief data and analytics officer Teresa Heitsenrether said employees are gaining one to two hours of productivity a week on average, while noting that the bank is still in the early stages.²¹

Most companies will not build their own platform at this scale, but the sequence works at any size: restrict while you assess, then replace the restriction with an approved path that people prefer.

Frameworks to borrow

You don’t have to design governance from scratch. The NIST AI Risk Management Framework is voluntary and organizes the work into four functions: govern, map, measure and manage, with governance designed to run through the other three.²² ISO/IEC 42001 is the international standard for AI management systems; it covers leadership, AI policy, risk management and monitoring, and organizations can choose to have it independently certified.²³ Use them as a starting structure, then adapt them to your company’s size and risk profile.

Shadow AI is what happens when employees adopt AI faster than leadership decides how it should be used. The companies that handle it well won’t be the ones with the strictest bans or the longest policy. They’ll be the ones where an executive owns the decisions, knows what’s running, matches oversight to risk, gives people approved tools they prefer, and reviews value and risk together every quarter. Your teams have already decided to use AI. The open question is whether leadership decides how.

If you’re mapping how AI is being used across your organization, shadow tools included, and want a practical, engineering-grounded view of where to put guardrails, that’s a conversation worth having early.