Shadow AI in the enterprise: how to find, govern and replace unsanctioned AI tools

Shadow AI in the enterprise: how to find, govern and replace unsanctioned AI tools

Shadow AI is the use of AI tools at work without IT or security approval. In Microsoft and LinkedIn’s 2024 survey, 78% of AI users brought their own tools to work (BYOAI), although the survey did not establish whether those tools were unauthorized. This guide shows how to find shadow AI, assess and govern its use, and provide approved alternatives where replacement is necessary.

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

Shadow AI is the use of AI tools, models or AI features inside an organization without approval from IT or security. Unlike classic shadow IT, the tool is often just a browser tab, and what leaves the company is data typed or pasted into a prompt. This guide covers employee use, not attacks on AI systems a company builds.

Shadow AI is a subset of shadow IT, but it can be harder to recognize because employees do not necessarily install new software or purchase a service.

DimensionShadow ITShadow AI
Typical formUnapproved applications, devices or cloud servicesAI chatbots, browser extensions, embedded assistants
Information exposedFiles, accounts and workloadsPrompts, source code, documents and business records
Data handlingExternal storage and processingExternal processing, with retention and model-use terms varying by provider
DiscoverySoftware inventory, purchases and loginsWeb traffic, personal accounts, integrations and enabled AI features
AdoptionOften initiated by a teamCan begin with an individual employee

Editorial comparison.

Three routes into the company: personal accounts, team-built integrations and AI inside approved software

Shadow AI commonly enters through three routes:

  1. Personal accounts: Employees use consumer AI tools for writing, analysis, research or coding.
  2. Team-built integrations: Departments connect an LLM API to an internal workflow without completing security review.
  3. Unreviewed or unmanaged AI functionality inside approved software: An existing SaaS application introduces AI capabilities that are enabled or used outside the organization’s established approval, visibility or governance scope.

The third route is easy to overlook. A company may have approved the original software without reviewing the data handling of its newly introduced AI features.

What this guide covers: employee and team use of AI tools, not attacks on AI systems you build

The focus here is unsanctioned employee and team activity, including browser extensions, API integrations and unmanaged agents.

Data leakage in this context results from how employees use tools. Attacks against an organization’s own LLM applications, such as prompt injection, are a separate AI security topic.

Why do employees use unapproved AI tools?

Employees may turn to unapproved AI tools for several reasons, including convenience, limited access to approved alternatives and unmet work needs. In Microsoft and LinkedIn’s 2024 survey, 78% of AI users brought their own tools to work (BYOAI). However, the survey did not establish whether those tools were unauthorized or identify the primary reasons for shadow AI use. The finding describes BYOAI behavior in 2024, not a measured shadow AI adoption rate for 2026.

Four reasons people go around IT, and what the data says about each

ReasonTypical situationOrganizational response
SpeedApproval takes longer than completing the taskSimplify low-risk approvals
Missing functionalityApproved software cannot perform the taskIdentify and prioritize the use case
ConvenienceA familiar personal tool is easier to accessImprove approved-tool usability
Limited awarenessEmployees do not understand the restrictionsProvide practical guidance and training

These are common explanations, not percentages established by the Microsoft and LinkedIn survey.

That survey found widespread bring-your-own-AI behavior, but it does not prove that every employee adopted unapproved tools for the same reason.

Microsoft

Reading shadow AI as a backlog of unmet needs

Shadow AI can reveal where official technology is failing to meet employee needs.

Repeated use of an unapproved document assistant, for example, may indicate demand for document summarization rather than deliberate resistance to security rules.

Instead of treating every discovery as an isolated violation, IT teams can group activity by business task and identify which capabilities deserve an approved alternative.

This approach aligns with MIT CISR’s 2026 research on minimum viable governance, which highlights the limitations of approval processes that are too slow for practical AI adoption.

MIT CISR

What does shadow AI put at risk?

Shadow AI puts company data, intellectual property, regulatory compliance and the reliability of business work at risk. The severity depends on what information employees share, how the provider processes it and whether the output influences consequential decisions. An unapproved tool is not automatically dangerous, but its risks have not been formally assessed.

Six risks, and what decides how serious each one is

RiskExampleWhat determines severity
Data exposureConfidential records pasted into an AI chatbotSensitivity, retention and provider controls
Intellectual propertyProprietary source code uploaded for debuggingOwnership, confidentiality and contractual terms
CompliancePersonal data processed without appropriate reviewApplicable laws and processing arrangements
Unreliable outputUnverified AI analysis used in a client reportImpact of errors and human review
Uncontrolled integrationsTeam API key accesses internal systemsPermissions and connected data
Poor accountabilityNo record of who approved an AI workflowOwnership, logging and auditability

Editorial risk assessment framework.

Why the same tool can be low risk on one plan and high risk on another

A product’s risk depends on its configuration and contractual conditions, not merely its brand.

A consumer account and an enterprise deployment of the same AI service may differ in authentication, administrative controls, data retention and contractual protections.

These differences must be verified against the relevant provider terms and actual configuration. Labels such as enterprise or business are not sufficient evidence by themselves.

How do you find shadow AI in your organization?

You find shadow AI by combining several discovery sources, because no single log shows every AI tool employees use. Network records reveal destinations, identity logs show connected applications, endpoint inventories expose extensions and local agents, and employee surveys reveal personal accounts. The objective is a reliable inventory, not perfect surveillance.

Six discovery sources, what each one finds and what it misses

Discovery sourceWhat it revealsWhat it may miss
Network and secure web gateway logsConnections to AI servicesActivity outside monitored networks
SaaS and identity logsAuthorized integrations and OAuth grantsPersonal accounts without corporate sign-in
Endpoint managementInstalled extensions, applications and agentsBrowser-only use on unmanaged devices
Expense and procurement recordsPaid AI subscriptionsFree accounts
API and cloud logsTeam-built model integrationsServices outside managed environments
Employee surveys and interviewsActual workflows and unmet needsForgotten or undisclosed usage

Editorial discovery matrix.

Microsoft’s 2026 documentation also describes network-based shadow AI discovery and detection of unmanaged agents, illustrating why discovery now needs to extend beyond traditional SaaS applications.

GitHub+1

Embedded AI and agents: tools nobody installed

An approved collaboration platform may activate a new AI assistant. A browser extension may gain AI functionality after an update. A developer may also run an agent locally without creating a conventional SaaS purchase record.

Discovery therefore needs to cover AI features and permissions inside existing software, not only unfamiliar application names.

Teams should review what these features can access, whether they transmit information externally and whether their use has an accountable owner.

A no-blame amnesty survey for what the logs cannot see

A short, non-punitive survey can help identify gaps in technical discovery.

Ask employees which AI tools they use, what tasks they perform, what information they share and which approved alternatives would meet their needs.

The purpose should be stated clearly: to understand usage and establish safer options, not to punish employees for disclosing it.

Survey results can be grouped by use case before individual applications are assessed.

What is an AI inventory, and what should it record?

An AI inventory is a maintained register of AI tools, models and AI-enabled features in use, each with an owner, a purpose and a description of the data it may access. NIST AI RMF Playbook control GOVERN 1.6 explicitly calls for mechanisms to inventory AI systems according to organizational risk priorities.

Nine fields that make an inventory entry usable

FieldWhat to record
Tool or featureApplication, model, agent or embedded capability
Business ownerPerson accountable for its use
Business purposeTask or workflow it supports
UsersTeams and roles with access
Data classificationInformation processed or shared
Provider and deploymentVendor, account type and hosting arrangement
Integrations and permissionsConnected systems and authorized actions
Decision and conditionsApproved, restricted or retired, with limitations
Review statusLast review, next review and responsible reviewer

Editorial inventory template informed by NIST GOVERN 1.6.

NIST describes inventories as a way to maintain visibility into deployed AI assets, relevant documentation and accountable parties.

AIRC

Discovery is a snapshot, the inventory is a process

A one-time discovery exercise becomes outdated as employees adopt new tools and vendors add AI functionality.

The inventory should therefore be updated through onboarding, procurement, integration reviews and periodic reassessment.

Every entry should have an owner who can confirm whether the application is still needed and whether its permissions remain appropriate.

How do you decide which shadow AI tools to sanction, restrict or retire?

Decide per tool, use case and configuration, not just per category. Sanction a tool when its provider terms, access controls and data handling meet your requirements, restrict it to lower-risk data when they only partly do, and retire it when the risks cannot be managed. The decision should account for the business need as well as security.

Five questions that put each tool into one of three outcomes

For each discovered tool, assess:

  1. What business task does it support?
  2. What data does it process?
  3. Where does that data go, and under which contractual conditions?
  4. Which identity, access and administrative controls are available?
  5. Is there a suitable approved alternative?

Then record one of three outcomes:

DecisionWhen it fitsRequired action
SanctionControls and terms meet organizational requirementsApprove, configure and add to the inventory
RestrictUseful tool with manageable limitationsDefine permitted data, users and tasks
RetireUnacceptable risk or a suitable approved replacement existsMigrate workflows and disable the old route

Editorial triage framework.

Why a blanket ban usually moves the problem out of sight

Blocking every AI service may reduce some visible usage without removing the underlying demand.

Employees may turn to personal devices, alternative services or unmonitored workflows. MIT CISR’s March 2026 research describes how overly burdensome governance can drive employees back toward unsanctioned tools.

MIT CISR

A more sustainable approach combines clear restrictions with an approved option that employees can access without unnecessary delays.

What should an AI acceptable use policy say, and why do policies alone fail?

An AI acceptable use policy tells employees which tools they may use, what information they may share and when approval is required. It should be short enough to apply during everyday work. A policy alone will not control shadow AI unless technical safeguards enforce it and employees have a practical approved alternative.

Eight sections of a one-page policy, with a data-class rule

A concise policy should cover:

SectionRule or information
1. ScopeWho and which AI activities are covered
2. Approved toolsWhere employees find permitted services
3. Data classificationWhich data categories may be submitted
4. Prohibited usesTasks or disclosures that are not allowed
5. Output reviewWhen human verification is required
6. New tool requestsHow employees request approval
7. Incident reportingHow to report accidental disclosures
8. Ownership and reviewWho maintains the policy and when it is updated

Editorial one-page policy outline.

A practical data-class rule could distinguish:

  • Public information: May be used in approved tools, subject to policy.
  • Internal information: Requires an approved service and appropriate safeguards.
  • Confidential, regulated or personal data: Requires explicit authorization and suitable processing controls.

The exact classification rules must follow the organization’s existing data policies.

Why a policy without enforcement and an alternative does not hold

Employees need more than a list of prohibited tools. They need to know how to complete legitimate work safely.

The policy should link to the approved tool catalogue, describe a quick request process and explain where employees can ask questions.

It should also be reinforced by identity controls, data loss prevention and periodic training.

An acceptable use policy governs day-to-day behavior. A broader enterprise AI governance approach addresses ownership, oversight and risk management across AI initiatives.

How do DLP and access control limit what employees can share with AI tools?

Data loss prevention (DLP) and access control reduce shadow AI exposure by limiting sensitive information sent to unapproved services. They work at different points, including endpoints, browsers, networks, identities and data repositories. No single layer sees every interaction, so effective protection combines technical controls with policy, monitoring and user education.

Seven control layers, and what each one cannot see

Control layerWhat it can restrictMain limitation
Identity and SSOAccess through managed identitiesPersonal accounts may bypass it
SaaS security controlsConnected applications and OAuth permissionsCoverage varies by integration
Secure web gatewayAccess to known AI destinationsUnmanaged networks may be invisible
Browser controlsUploads and interactions in managed browsersOther browsers and devices
Endpoint DLPSensitive files and clipboard activityUnmanaged endpoints
Data classification and repository permissionsAccess to protected informationCannot control every downstream use
API gateways and cloud controlsTeam-built model integrationsDoes not cover independent consumer use

Editorial control-layer matrix.

DLP should use existing data classification rules and focus on consequential exposure, rather than indiscriminately inspecting every employee interaction.

Controls also need to be tested against actual workflows. An overly broad restriction can block legitimate work while leaving other routes unprotected.

Where employee-side controls end and controls inside your own AI systems begin

Shadow AI controls focus on what employees may send to external or unapproved AI tools.

Security controls inside an organization’s own LLM applications address a different boundary: what the model may retrieve, generate or execute.

For example, blocking confidential uploads to an external chatbot is different from enforcing document-level authorization in an internal assistant.

Both matter, but they should be assigned to the appropriate systems and owners. Enterprise integrations also need API-level access controls.

How do you replace unsanctioned AI tools with an approved AI platform?

Replacing shadow AI requires more than purchasing an enterprise AI license. First identify the work employees are doing, then provide an approved route that meets the need with appropriate data access, security and support. Migrate the workflows, verify adoption and retire the unsanctioned route only when the replacement is usable.

Six steps from use-case ranking to retiring the old route

  1. Rank the use cases. Group discovered activity by business value, frequency and data sensitivity.
  2. Define requirements. Specify necessary features, integrations, permissions and output quality.
  3. Select the replacement. Decide whether to configure an existing platform, purchase a solution or build an application.
  4. Pilot with real users. Test representative tasks, security controls and usability before broad rollout.
  5. Migrate and train. Provide templates, instructions, access and support for the approved workflow.
  6. Retire and verify. Remove unnecessary permissions, restrict the old route and monitor whether employees actually switch.

The sequence matters. Retiring a familiar tool before its replacement works can recreate the original shadow AI problem.

Configure, buy or build: which route fits which situation

RouteBest fitMain consideration
ConfigureCommon tasks supported by existing enterprise softwareFast adoption with existing controls
BuySpecialized capability available from a suitable providerVendor review and integration
BuildDifferentiating workflow requiring custom data and logicEngineering and operational ownership

An approved enterprise assistant may be enough for general writing and summarization. A custom knowledge assistant may need retrieval-augmented generation and integration with internal documents.

The right solution is the one that addresses the actual employee task while meeting the organization’s requirements.

When the shadow tool is something employees built

Some shadow AI tools are not consumer applications. They are internal scripts, automations or assistants created by business teams.

These should be assessed for dependencies, credentials, ownership and maintenance before they are sanctioned or replaced.

The broader practice of governing employee-built applications belongs to citizen development. Shadow AI triage should identify such solutions and route them into the appropriate engineering review.

What the platform stands on: data, access and operations

An approved AI platform needs more than a user interface.

Its underlying AI-grade foundation should support reliable data access, identity and permission enforcement, monitored integrations, evaluation and operational ownership.

For more complex use cases, the platform may also support enterprise LLM applications and AI agents connected to business systems.

An approved platform succeeds when employees can complete their work through a route that is both useful and governed.

How do you keep shadow AI from coming back?

Keep shadow AI from returning by giving each control a named owner, repeating discovery on a schedule, training employees on the approved route and tracking adoption. The share of AI usage going through approved tools is especially important because it shows whether the replacement is working, rather than merely whether restrictions have been introduced.

One named owner, one cadence and one metric per control

ActivityTypical ownerReview cadenceMetric
Discover AI usageIT and securityMonthly or continuousUnreviewed tools discovered
Maintain inventoryAI governance ownerMonthlyEntries with current reviews
Assess new requestsIT, security and business ownerPer requestTime to approval decision
Enforce data controlsSecurityContinuousPolicy violations requiring action
Migrate workflowsPlatform and business teamsPer rolloutUse cases moved to approved tools
Train employeesIT enablement and business teamsPeriodicTraining completion and feedback
Review embedded AISaaS application ownersOn major changesAI features reassessed

Illustrative operating model. Cadences should reflect organizational risk and scale.

The program should also track the number of tools retired and the proportion of observed AI usage occurring through approved services.

These are management indicators, not proof that all unsanctioned activity has been eliminated.

Recertifying AI features inside approved software

An application approved last year may introduce a new assistant, agent or model integration this year.

Software owners should reassess material changes to AI functionality, provider terms, data handling and permissions.

This prevents the AI inventory from becoming a list of products that were once approved but are no longer understood.

How does a shadow AI program fit into AI governance, responsible AI and the EU AI Act?

A shadow AI program is one operational component of AI governance: it discovers what is in use, applies the policy and feeds the inventory that governance frameworks expect. It does not replace those frameworks, responsible AI practice or EU AI Act compliance, which sit above it and require separate assessment.

Five disciplines, five different questions

DisciplineMain question
Shadow AI managementWhich unapproved AI tools are in use, and what should happen to them?
AI governanceWho approves, owns and oversees AI systems?
Responsible AIAre AI systems used appropriately and with consideration for potential harm?
AI securityCan systems, data or connected tools be compromised or misused?
Legal complianceWhich obligations apply to the specific AI system and its use?

Editorial comparison.

NIST’s AI Risk Management Framework provides broader guidance on managing AI risk. Its GOVERN 1.6 control supports maintaining an inventory, but an inventory alone does not establish responsible AI or legal compliance.

Similarly, the EU AI Act requires assessment of applicable obligations based on the AI system and the organization’s role.

Article 4 also requires AI providers and deployers to take measures supporting the development of AI literacy among staff and others using AI systems on their behalf. These measures should reflect the users’ knowledge, experience and the context in which AI is used.

The obligation has applied since 2 February 2025, with supervision and enforcement beginning on 2 August 2026. Following amendments introduced in July 2026, Article 4 does not require organizations to guarantee a specific level of AI literacy for each individual.

For a shadow AI program, this makes employee guidance and training particularly relevant. Organizations can help employees understand which AI tools are approved, what information they may share, how to recognize risks and how to request approval for new AI uses.

Finding an unapproved AI tool is the beginning of assessing applicable obligations, not its conclusion.

The practical goal is to connect discovery, approval, data protection, employee education and approved alternatives where necessary to the organization’s wider governance process.

FAQ

Is shadow AI always bad?

No. Employees may use unapproved AI tools to solve legitimate business problems. The issue is that the organization has not assessed the tools, their data handling or their permissions. Some can be approved or restricted, while others should be retired.

What is shadow AI in cyber security?

Shadow AI refers to AI tools, features or agents used without appropriate organizational approval or visibility. In cybersecurity, the concern is that these tools may expose data, bypass controls or create unmanaged integrations.

Can shadow AI exist if nobody installs anything?

Yes. Employees can access AI through a browser, personal account or an AI feature added to already approved software. Teams can also connect to model APIs without installing a conventional application.

What is the difference between an AI acceptable use policy and an AI governance policy?

An AI acceptable use policy explains what employees may do with AI tools and data. An AI governance policy establishes broader responsibilities, oversight, risk management and approval processes for AI across the organization.

What is BYOAI?

BYOAI means Bring Your Own AI. It describes employees bringing personally selected AI tools into their work. Microsoft and LinkedIn’s 2024 Work Trend Index reported that 78% of surveyed AI users did so.

Sources

https://www.microsoft.com/en-us/worklab/work-trend-index/ai-at-work-is-here-now-comes-the-hard-part

https://news.microsoft.com/source/2024/05/08/microsoft-and-linkedin-release-the-2024-work-trend-index-on-the-state-of-ai-at-work

https://cisr.mit.edu/publication/2026_0301_GenAIGovernance_VanderMeulenJewerLevallet

https://cisr.mit.edu/publication/2026_0101_GenerativeAIRisk_VanderMeulenLefebvreWixom

https://airc.nist.gov/airmf-resources/playbook/govern

https://learn.microsoft.com/microsoft-365/admin/manage/agent-shadow-ai

https://github.com/MicrosoftDocs/entra-docs/blob/main/docs/global-secure-access/concept-shadow-ai-discovery.m

https://www.nist.gov/itl/ai-risk-management-framework

Translate »