What Is Shadow AI? How to Govern AI Use You Cannot Yet See

Shadow AI is not simply employees breaking AI policy. It is a visibility and operating-model gap that can hide sensitive data use, unassessed suppliers, connected agents and genuine unmet demand.

Decision brief

In brief

Shadow AI is AI used for organisational work outside sufficient visibility, approval or governance. KOR treats it as both a control exposure and an adoption signal, using a five-stage control loop — Discover, Classify, Provide, Control and Review — to move material AI use into proportionate, governed adoption.

Signal
Employees and teams are using personal, free, departmental or embedded AI outside normal approval and governance routes, while management reporting still reflects only centrally licensed or registered systems.
Implication
The official AI inventory can materially understate both organisational exposure and genuine employee demand, leaving leaders unable to distinguish harmless experimentation from sensitive data processing, consequential decisions or connected agents.
Decision
Build a living view of real AI use, classify material use cases by data, decision and action risk, provide credible approved routes for legitimate demand, and apply proportionate controls rather than relying on blanket prohibition.

An organisation can have an AI policy, an approved enterprise assistant and a steering committee — and still have no reliable picture of how artificial intelligence is actually being used across the business.

An employee summarises a confidential document using a personal AI account. A team quietly subscribes to an AI meeting assistant. A developer connects an AI coding tool to an internal repository. A manager starts using an AI feature embedded inside an existing SaaS platform without anyone reassessing what happens to organisational data.

None of those activities necessarily appears on the official AI inventory.

This is Shadow AI.

The UK's National Cyber Security Centre defines Shadow AI as AI technology used outside an organisation's approved systems and processes. It is a form of Shadow IT, but the operating implications are broader because AI systems can analyse information, generate outputs, influence decisions and, increasingly, take actions through connected tools and agents. The NCSC's September 2026 guidance also makes an important practical point: organisations should focus on reducing the risk rather than assume Shadow AI can be eliminated completely.

The scale is no longer theoretical. Deloitte UK's 2026 GenAI Workforce Survey, based on 25,000 UK workers, found that 63% had knowingly used generative AI for work. Among GenAI users, 31% said they used it without their employer's knowledge, 17% paid for at least one work-related GenAI tool themselves and 46% used free tools at work.

Our central argument is straightforward:

Shadow AI is not primarily an employee misconduct problem. It is a visibility and operating-model gap.

Unmanaged use can create real exposure. It can also reveal where employees have found useful capability before the organisation has provided a governed route to it.

The leadership task is therefore to establish where Shadow AI is happening, why it exists, what data and permissions are involved, and which uses should be stopped, constrained, replaced or brought into the formal AI estate.

In brief

Signal: AI adoption is extending beyond centrally approved tools, licences and formal programmes.

Implication: The official AI inventory can understate both organisational exposure and genuine employee demand.

Decision: Discover the real environment first. Then classify use cases according to the information involved, the decision being influenced and the actions the AI can perform.

What Shadow AI means in practice

Shadow AI is the use of artificial intelligence within or for an organisation without sufficient visibility, approval or governance through the organisation's normal technology, security, procurement, data or AI-management processes.

The important condition is not simply whether a product appears on an approved list. It is whether the organisation understands and controls the relevant use.

An employee may use a public chatbot that IT has never seen. That is the obvious case. Shadow AI can also appear inside technology the organisation already owns.

A SaaS supplier may introduce a generative feature after the original supplier assessment. A department may buy an AI application from its own budget. A developer may connect an approved model to an unapproved tool or dataset. An employee may use an approved enterprise assistant for information or decisions that sit outside the permitted use case.

The AI product can therefore be reputable while the use case, data flow, integration, account configuration or permission model remains unmanaged.

That is why an approved-application list is useful but insufficient.

ExampleShadow AI?Reason
Employee uses an approved enterprise AI assistant for an authorised use caseGenerally noThe tool, account and use sit within organisational controls
Employee pastes a confidential supplier contract into a personal AI accountYesOrganisational information has entered an unmanaged service
Marketing purchases an AI content platform without technology or procurement reviewYesThe supplier and data use sit outside the approved process
An existing SaaS supplier activates a new AI feature that has never been assessedPotentiallyThe supplier may be approved while the new AI processing is not
A developer connects an AI assistant to an internal code repository without registering the integrationYesNew data access and technical permissions exist outside governance
An approved enterprise AI tool is deliberately used in a prohibited high-risk workflowThe tool is approved; the use is notThe governance gap exists at use-case level rather than product level
A business team creates an internally hosted AI agent without central registrationYesShadow AI does not have to involve a public or consumer AI service

This distinction becomes more important as AI moves from standalone chat interfaces into existing applications, APIs, browsers, workflow tools and autonomous agents.

The emerging visibility problem is therefore broader than identifying which chatbot employees use. Organisations increasingly need to understand which models, agents, connectors and AI-enabled features can access which information, systems and actions — and who is accountable for that access.

Shadow AI, Shadow IT and BYOAI are related but different

Shadow AI developed from a familiar technology-management problem: Shadow IT.

Employees have always found their own software when official technology did not meet their needs. Consumer file-sharing services, messaging applications, spreadsheets, browser extensions and departmental SaaS subscriptions created similar visibility problems.

AI changes the control boundary.

Traditional Shadow IT may store organisational information or create an unmanaged application. An AI system may additionally analyse that information, combine it with other context, generate content used in business decisions, produce code, interact with external services or initiate actions through an agent.

Bring Your Own AI, or BYOAI, is one manifestation of Shadow AI: employees bringing personally selected or personally funded AI tools into their work.

Shadow AI is broader. An internally developed agent can be Shadow AI. An AI feature inside an approved supplier can become Shadow AI. An unmanaged API integration can be Shadow AI.

The common condition is insufficient organisational visibility or control over AI capability being used to perform organisational work.

Shadow AI often reflects unmet demand

It is easy to frame Shadow AI as people ignoring policy. Sometimes that will be accurate. It is rarely the complete explanation.

Employees use technology because they have work to complete. Generative AI can reduce the effort involved in research, summarisation, drafting, analysis, coding and information retrieval. If employees find useful capability before the organisation provides an approved equivalent, an incentive to bypass the formal route already exists.

Deloitte's 2026 research supports that interpretation. Among workers paying for their own GenAI or using a tool that might not be approved, 21% said those tools performed better than company alternatives and 14% said the technology was essential to their job but not funded by their employer. Deloitte's Shadow AI analysis argues that stricter rules alone are unlikely to solve the problem.

That makes Shadow AI an adoption signal as well as a control signal.

This aligns with KOR's broader Enterprise AI Adoption Model: access, activity, adoption and realised value are different findings. An organisation can restrict access while demand continues. It can purchase licences without changing behaviour. It can also have widespread informal activity without understanding whether that activity is safe, sustained or valuable.

Shadow AI sits inside that gap.

The KOR Shadow AI Control Gap

KOR proposes a four-layer view of Shadow AI. The purpose is not to rank every use on a maturity scale. It is to identify where visibility and control can fail as AI becomes more embedded and more capable.

Framework boundary: The KOR Shadow AI Control Gap is a proposed operating model, not an industry standard or externally validated risk taxonomy. It should be adapted to the organisation's technology, regulatory exposure and risk appetite.

Layer 1: personal AI

This is the most visible form of Shadow AI.

Employees use personal or free accounts for generative AI applications, image generators, research assistants, writing tools, transcription services or coding assistants.

The product may be highly capable. The organisational problem is that the account configuration, contractual terms, retention settings, data processing and user behaviour may sit outside enterprise control.

Layer 2: departmental and embedded AI

A function independently purchases an AI-enabled application, activates a trial or starts using an AI capability already included within a SaaS product.

This layer is easy to miss because the underlying supplier may already exist in the organisation.

The supplier approval therefore needs to be distinguished from approval of the new AI capability. KOR's AI supplier due-diligence framework applies the same principle across discovery, evaluation, contracting, monitoring and exit: supplier assurance has to follow the live capability, not stop at the original contract.

Layer 3: developer and automation AI

Developers, analysts and technically capable business users can assemble AI workflows rapidly using APIs, low-code tools, scripts and model integrations.

These uses can create material value. They can also create connections between models and internal data that do not appear in a conventional application inventory.

The relevant unit of control is therefore often model + data + integration + permission, not simply the name of the AI provider.

Layer 4: agents and connected AI

The control exposure increases when AI can take action.

An isolated chatbot with no internal permissions creates a different risk profile from an agent that can read files, access email, query databases, execute code, create tickets or change enterprise records.

The NCSC's Shadow AI guidance highlights this issue directly: AI agents can inherit access to data, services and privileges, meaning a vulnerability or misconfiguration can expose more than the agent itself.

Organisations therefore need visibility over both which AI systems are present and what they are authorised to do.

The risks are not equal

Not every instance of Shadow AI presents the same level of exposure.

Using a personal AI account to rewrite a generic meeting agenda is not equivalent to uploading employee records, reviewing acquisition documents or connecting an agent to a finance system.

Effective governance starts by separating those situations.

Data visibility and confidentiality

The most immediate concern is organisational information leaving a governed environment.

Commercial contracts, customer information, personal data, source code, financial information, intellectual property or internal strategy may be submitted to services without the organisation fully understanding how that information is processed, retained or accessed.

The NCSC warns that transferring sensitive or proprietary information to consumer AI services can reduce organisational visibility and control, particularly where data may be stored, retained or used to improve a service outside established arrangements.

Data protection and regulatory accountability

Calling an application Shadow AI does not remove organisational accountability.

Where personal data is processed, the relevant data-protection obligations still need to be considered. The ICO's AI contracts and third-parties guidance expects organisations to understand roles across the processing chain, undertake due diligence, document technical and organisational controls and monitor continuing supplier compliance. The ICO currently notes that parts of this guidance are under review following the Data (Use and Access) Act.

An organisation that does not know processing is taking place is poorly placed to assess that processing.

Visibility is therefore a precondition for meaningful governance.

Intellectual property and contractual exposure

Employees may expose confidential material to AI services without knowing whether the applicable account, licence or contract provides the protections expected by the organisation.

The reverse problem also exists. Generated material may be incorporated into organisational work without sufficient understanding of provenance, permitted use or reliability.

These issues are not resolved by a generic AI policy. They require appropriate supplier assurance, contractual treatment and practical user guidance.

Decision quality and human oversight

Shadow AI can influence visible decisions while remaining invisible itself.

A sourcing recommendation, candidate summary, financial analysis, supplier assessment or management paper may be partly generated by an AI system without the organisation recording that fact.

The risk is not that every AI-generated output is wrong. The risk is that the organisation has not defined a review standard proportionate to the decision.

This matters particularly in people decisions. KOR's AI Adoption in HR analysis separates lower-risk assistance from recommendation and consequential decision support because the appropriate human-oversight boundary changes with the use case.

Security and system access

Risk increases sharply when AI can access organisational systems.

An isolated chatbot with no internal permissions and an agent capable of querying customer data and taking actions should not sit in the same control category.

Governance needs to assess capability and access, not merely application name.

Fragmented spend

Shadow AI is also a procurement problem.

Employees may independently pay for overlapping products. Teams may acquire tools on cards or expense claims. Functions may contract with suppliers without consolidating demand.

The result can be duplicated spend alongside weaker enterprise terms, limited administrative visibility and reduced commercial leverage.

Lost organisational learning

There is an opportunity cost as well as a control risk.

If useful AI experimentation happens invisibly, the organisation cannot easily identify which workflows employees are improving, which products genuinely work, what adoption barriers exist or where investment could create value.

Shadow AI can therefore conceal innovation as well as exposure.

That is one reason discovery should not begin as a disciplinary exercise.

A policy cannot govern what the organisation cannot see

The NCSC's current guidance is pragmatic: Shadow AI is unlikely to disappear completely, so organisations should reduce risk rather than assume it can simply be eliminated.

The UK Government's AI Management Essentials reinforces the underlying management requirement. It asks organisations to maintain a complete and up-to-date record of AI systems they develop and use, establish a process for adding new systems, request relevant documentation from third-party providers and review the record regularly.

A register alone will not discover everything. It does establish the organisational expectation that AI use is a managed system rather than an informal collection of tools.

KOR's interpretation is that the governed route needs to become easier than the shadow route.

If obtaining an approved AI capability takes months while a useful public service takes seconds to open, policy is competing against convenience. Some employees will route around the policy.

Governance needs to change that operating equation.

The KOR Shadow AI Control Loop

KOR proposes a five-stage operating cycle for moving from invisible AI activity towards governed adoption.

1. Discover

Start with the real environment rather than the official inventory.

Discovery can combine employee surveys and interviews with procurement records, payment-card and expense data, software inventories, browser or endpoint visibility, identity logs, network or cloud controls and existing security tooling.

No single source will reveal everything.

The objective is not merely to list AI products. It is to understand who is using AI, for what work, with what data, through which accounts and with what level of system access.

The NCSC summarises the principle simply: organisations cannot manage what they do not know about.

2. Classify

Discovery should not produce one blanket Shadow AI risk label.

Material use cases should be classified across four dimensions:

  • Data sensitivity — public information, internal business information, confidential data, personal data, source code or regulated information.
  • AI authority — drafting, analysis, recommendation, decision support or autonomous action.
  • Operational criticality — optional individual productivity through to core process or regulated decision.
  • Control environment — account type, enterprise administration, contractual protections, retention settings, logging, identity and integration controls.

This produces a more useful risk picture than dividing the world into approved AI and banned AI.

3. Provide

Where Shadow AI reveals a legitimate business requirement, the organisation should decide whether a governed route can be provided.

That may mean enabling an existing enterprise assistant, procuring a specialist application, giving teams controlled access to an approved model environment or creating a defined experimental route for new use cases.

Not every request should be funded.

Repeated Shadow AI activity around the same workflow is still useful evidence that a capability gap may exist.

4. Control

Technical and organisational controls should match the risk.

Low-risk experimentation may require clear rules and practical education. Sensitive data may require approved enterprise accounts and data-loss controls. Agentic systems may require restricted permissions, logging, sandboxing, human approval and explicit action boundaries.

The principle is proportional control.

If every use is treated as high risk, governance becomes difficult to use and employees find routes around it. If every use is treated as low risk, the organisation has not governed anything material.

5. Review

Shadow AI is not a one-off remediation exercise.

AI functionality is being added continuously to software organisations already own. Employees discover new products. Existing products change their terms. Agents gain additional integrations. Business processes evolve.

The inventory therefore needs to remain a living view of organisational AI rather than an annual compliance spreadsheet.

Discovery, classification and review should become part of the AI operating model.

Shadow AI requires shared ownership

Shadow AI should not be assigned to IT alone.

Technology and security teams are essential, but the problem crosses organisational boundaries.

Security can identify technical exposure and access. Data-protection and legal teams can assess processing and legal risk. Procurement can identify suppliers, subscriptions, duplicated spend and contractual gaps. HR and learning teams influence acceptable use, capability and employee behaviour. AI or transformation leadership can decide which experiments should become enterprise capability. Business functions understand why employees are using the technology in the first place.

No function has the complete picture individually.

Shadow AI therefore exposes a broader operating-model issue: whether AI discovery, supplier governance, data protection, security, employee adoption and business value are connected through clear ownership and escalation.

Where those responsibilities remain fragmented, Shadow AI is likely to remain fragmented too.

Procurement has a wider role than blocking purchases

Shadow AI is particularly relevant to procurement because the function sits across both sides of the problem.

It can help discover AI entering the organisation through supplier contracts, departmental purchases, expenses and SaaS renewals. It can also help convert repeated informal demand into appropriate enterprise agreements.

The same distinction appears in KOR's Procurement AI Readiness work: visible AI activity does not prove that the organisation has the data, governance, workflow integration, supplier controls, adoption and evidence required for controlled scale.

Supplier questions also need to become more precise:

  • Does the product now contain AI functionality?
  • Which models or subprocessors are involved?
  • What organisational information can the system process?
  • Is customer data used to train or improve models?
  • What retention and deletion controls exist?
  • Can AI features be disabled?
  • What administrative visibility and usage evidence are available?
  • What happens when the supplier changes models or introduces new AI functionality?

Those questions increasingly belong inside mainstream supplier governance rather than a separate AI-project process.

The same applies to existing suppliers. An organisation may have assessed a SaaS service several years ago and still know very little about the AI capability now embedded inside it.

Procurement also has an adoption role. KOR's Procurement AI Adoption Model treats approved use as an operating capability: the accessible, useful route should also be the governed route. Shadow AI is a practical test of whether that condition exists.

The goal is not zero Shadow AI

A zero-Shadow-AI target sounds reassuring. It is unlikely to be a useful operating objective.

AI capability is being embedded into browsers, productivity suites, search tools, enterprise SaaS, development environments and consumer applications. The boundary between an AI application and ordinary software is becoming less stable.

A more useful objective is minimum unmanaged material AI risk.

That means the organisation can identify consequential AI use, understand the data and permissions involved, provide credible approved routes for legitimate needs and intervene proportionately where activity creates unacceptable exposure.

Some Shadow AI should be stopped.

Some should be constrained.

Some should be replaced.

Some will turn out to be useful experimentation that should be formalised and scaled.

From Shadow AI to governed adoption

Shadow AI is often discussed as though it sits outside AI strategy. In practice, it tells leaders something important about that strategy.

It shows where employees believe AI can help. It identifies workflows where official capability has not caught up with demand. It reveals gaps in supplier governance and technology visibility. It also tests whether AI governance works when confronted with real employee behaviour.

That makes Shadow AI both a risk indicator and an adoption signal.

Organisations that respond only by banning applications may reduce some immediate exposure while learning very little about the unmet need that caused the behaviour. Organisations that ignore Shadow AI get the opposite problem: experimentation continues without adequate visibility or control.

The stronger operating model sits between those extremes.

Discover the real environment. Classify material use. Provide governed routes where they make sense. Apply proportionate controls. Review continuously as systems and behaviour change.

That is the transition from Shadow AI to governed adoption.

How KOR Intelligence approaches Shadow AI

The first useful output is not another generic AI policy. It is a credible map of the current environment: the AI tools and capabilities already being used, the business needs driving them, the information they can access, the suppliers involved, the controls already in place and the gaps that require action.

That evidence supports a more disciplined decision: what should be prohibited, what can be tolerated, what needs stronger control and what should be brought into the formal AI estate.

KOR approaches this through the wider relationship between AI discovery, governed adoption, supplier assurance and operating-model design. Our AI discovery and governance work is intended to help organisations connect those questions rather than manage them as separate technology, procurement and risk exercises.

The practical objective is not maximum control or maximum AI use. It is operational clarity over where AI is creating value, where it is creating exposure and which action the organisation should take next.

References

  1. 01The NCSC's September 2026 guidance · Research source · accessed 2026-09-19
  2. 02Deloitte UK's 2026 GenAI Workforce Survey · Research source · accessed 2026-09-19
  3. 03Deloitte's Shadow AI analysis · Research source · accessed 2026-09-19
  4. 04ICO's AI contracts and third-parties guidance · Research source · accessed 2026-09-19
  5. 05AI Management Essentials · Research source · accessed 2026-09-19