What is shadow AI?

By Identra · Updated October 3, 2026

Shadow AI is the use of AI tools, models, or agents for work outside an organization’s approved governance and oversight. It includes unapproved services and unreviewed uses of approved tools, such as connecting a personal AI account to company files.

Key numbers

78% of AI users bring their own AI tools to work (Microsoft and LinkedIn Work Trend Index, 2024)
20% of studied organizations suffered a breach linked to shadow AI (IBM Cost of a Data Breach Report, 2025)
$670,000 added to the average breach cost when shadow AI was involved (IBM Cost of a Data Breach Report, 2025)
42% of generative AI data policy violations involved source code (Netskope Cloud and Threat Report, 2026)

Why does shadow AI happen?

Employees may adopt AI to summarize documents, write code, prepare presentations, or automate repetitive work. Shadow AI emerges when that adoption bypasses the organization’s review process. An unclear policy, a slow approval process, or an approved tool that cannot support the task can contribute.

The gap is about the use as well as the product. A company might approve a chatbot for public marketing copy without approving customer records, repository access, or autonomous actions. Effective AI governance defines which accounts, data, integrations, and tasks an approval covers.

Shadow AI does not require malicious intent or a personal account. A team can buy a service with a company card and still leave security unaware of its data access, owner, or retention settings.

How is shadow AI different from shadow IT?

Shadow AI is a form of shadow IT focused on AI capabilities. The categories overlap, but AI review needs to examine submitted content, generated output, and any authority delegated to agents. A familiar application can also introduce an AI feature that needs its own review.

Approval should have a defined scope. Calling a vendor approved does not establish that every account, connector, or agent built with its tools is governed.

Shadow IT
Enterprise example
A team adopts an unapproved file-sharing service
Review focus
Ownership, access, data handling, and lifecycle
Shadow AI
Enterprise example
An employee connects an unreviewed AI assistant to company files
Review focus
Data handling, permissions, outputs, and possible actions
Governed AI use
Enterprise example
A managed assistant operates within an approved use case
Review focus
Whether accounts, data, and permissions remain within approval

Category	Enterprise example	Review focus
Shadow IT	A team adopts an unapproved file-sharing service	Ownership, access, data handling, and lifecycle
Shadow AI	An employee connects an unreviewed AI assistant to company files	Data handling, permissions, outputs, and possible actions
Governed AI use	A managed assistant operates within an approved use case	Whether accounts, data, and permissions remain within approval

What risks does shadow AI create?

Sensitive information can reach a service without the organization reviewing its handling terms or settings. AI data leakage can involve prompts, uploaded files, meeting transcripts, or connected repositories. Whether a provider retains content or uses it for training depends on the service, account, contract, and configuration. Neither outcome should be assumed.

Connected tools introduce a separate access risk. An assistant with permission to read files can reach information beyond what an employee manually pastes. An agent with write permissions may change records or run commands. Reviewing OAuth app risk helps establish what a connected application is authorized to reach.

Unreviewed outputs can also introduce errors into code, customer communications, and business decisions. Missing ownership or accessible activity records can make an incident harder to investigate. These are possible consequences of weak oversight, not proof that every unapproved tool has exposed data.

What does shadow AI look like in practice?

Consider a hypothetical sales team that adopts an AI meeting assistant without review. An employee signs up with a personal email address, uploads a customer-call recording, and connects a work calendar to simplify scheduling. The recording includes confidential pricing and customer information.

Security later discovers the subscription during an expense review. The immediate questions are concrete: Which recordings were uploaded? What calendar permissions were granted? Who can access the transcripts? Can the company obtain activity records and request deletion?

The response should separate confirmed exposure from possible access. An upload establishes that a particular file was shared. A permission grant establishes authorized reach, not that every reachable item was read. The team should pause further sharing, preserve available evidence, review and revoke unnecessary access, and move the workflow to an approved service.

Offboarding should explicitly address connected apps and credentials. Whether access persists after account disablement depends on the provider and authorization model, so responders should verify revocation rather than assume either continued access or complete removal.

How do you discover shadow AI?

Combine employee reporting with technical evidence. Network records can reveal contact with an AI service, but a domain alone usually cannot establish the signed-in account, submitted content, or business purpose. Account-aware AI access matters when personal and company accounts use the same service.

Use discovery to build an actionable inventory. For each finding, record the business owner, account type, use case, data involved, connected permissions, and approval status. Mark unknowns explicitly.

Review identity-provider app registrations, consent grants, and available sign-in records.
Inspect managed endpoints for AI clients, coding assistants, local models, and agent tooling.
Review browser extensions and available browser or network activity within established employee privacy rules.
Check procurement and expense records for subscriptions outside the approved catalog.
Ask teams which AI features and workflows they use, including features embedded in existing SaaS products.

How do you bring shadow AI under control?

Prioritize findings by sensitive data exposure and action permissions. An agent that can modify production resources needs different controls from a drafting tool limited to public material. Apply least privilege to connectors, service accounts, and agent permissions.

Publish an AI acceptable use policy with recognizable examples and a clear review route. Blocking is appropriate for prohibited activity. Approved alternatives and timely reviews help employees complete legitimate work within policy.

Define allowed tools, account types, data classes, and tasks. State which changes require another review.
Provide managed accounts and verify the service’s retention, training, sharing, and administrative settings.
Remove unnecessary connectors and permissions. Assign an owner to each integration and credential.
Require human review for consequential outputs and approval before high-impact agent actions.
Document how to report accidental sharing, preserve evidence, revoke access, and request deletion.
Repeat discovery and review when tools, permissions, owners, or use cases change. Include AI access in offboarding.

How Identra thinks about it

Identra helps organizations make shadow AI visible across the browser, endpoint, and identity, SaaS, and cloud environments. Teams can guide supported browser AI use toward company accounts, apply policies to sensitive sharing and agent actions, and revoke risky connected-app grants.

Go deeper: AI discovery across identity, SaaS and cloud

Frequently asked questions

Is using a personal AI account always shadow AI?

It is shadow AI when the workplace use falls outside approved governance. An organization could explicitly allow a limited personal-account use case, but that permission should specify acceptable data and tasks.

Can an approved AI product still be used as shadow AI?

Yes. An employee might use an unapproved account, enable an unreviewed connector, or submit data outside the approved use case. Approval needs to cover the configuration and activity.

Are local AI models considered shadow AI?

They can be if used for work without the required oversight. Local processing may reduce some external sharing risks, but file access, stored outputs, installed software, and agent actions still need review.

Does shadow AI automatically mean a data breach occurred?

No. It establishes a governance gap, not a confirmed breach. Investigators need evidence of what was shared, accessed, or changed before determining impact.

Can blocking AI websites eliminate shadow AI?

Website blocking can restrict known destinations, but discovery also needs to cover desktop clients, API integrations, local tools, and AI features within approved applications. Controls should match the actual workflow.

Compare

All comparisons →