Blog
AI Governance
AI Security
Find shadow AI across chatbots, SaaS features, APIs, and local agents. Build an evidence-based inventory and identify the uses that need review first.
An employee can use an approved AI product through a personal account. An approved SaaS platform can add a model-backed feature. An internal script can call a model API without creating a purchase order. A vendor list can miss all three.
Shadow AI is AI use outside an organization’s approved systems or processes. Discovery means finding those uses, establishing what the evidence shows, and assigning someone to resolve the unknowns. A list of AI domains is a starting point.
The UK National Cyber Security Centre’s September 2026 guidance connects this problem to a practical gap: employees need tools that help them do their work, while organizational policies and approved services may not keep pace. Understanding that work helps a review team offer an alternative people can use.
For GRC teams, the useful output is an inventory of AI use with evidence, owners and review status. This guide explains where to look and how to avoid treating a weak signal as a confirmed data exposure.
Where does shadow AI hide?
Public chatbots are one place to look. Discovery also needs to cover AI inside existing applications and software running on employee devices or internal infrastructure.
Shadow AI usually shows up in three forms:
Standalone AI services: chatbots, writing tools, meeting assistants and other applications accessed through personal or managed accounts.
AI inside approved software: assistants, summarization features and agents enabled within a SaaS product the organization already uses.
Local and internally built AI: desktop tools, browser extensions, coding agents, scripts, model runtimes and applications that call model APIs.
An embedded feature may not produce a new vendor record or an obvious new domain in employee traffic. Establish which features are enabled and where the supplier processes data. Do not assume that every AI feature sends data to an additional provider.
Why is shadow AI a risk for GRC teams?
An unknown use leaves unanswered questions about data, permissions and accountability. Some existing controls may still apply. Discovery should establish their coverage, rather than assume the tool is either completely uncontrolled or safe because it runs on a managed device.
Here are the main risks:
Data handling: identify which information was uploaded, retrieved or generated, and which account and terms applied. A connection to an AI service alone does not establish that sensitive data was sent.
Unreviewed dependencies: distinguish the publisher, operator and recipients of data. An internal application or local model is not automatically a new vendor, but it still needs an owner and a security review appropriate to its use.
Decision and action risk: a drafting tool, a hiring recommendation and an agent that changes access rights create different consequences. Record what people or systems do with the output.
Evidence gaps: unknown accounts or undocumented features can make it difficult to establish what was approved and which controls applied. Keep uncertainty visible until the owner supplies evidence.
NIST’s voluntary Generative AI Profile includes inventories of third parties with access to organizational content and due diligence for embedded AI. That supports treating discovery as an ongoing governance activity, with a clear handoff to technical and supplier review.
How to find shadow AI: six sources of evidence
Each source answers a different question. Record which users, devices, networks and applications it covers, then combine the results. Missing activity can mean there was none, or that the activity took place outside the collection boundary.
Review network and application discovery. Use authorized DNS, proxy, secure web gateway and SaaS discovery records to identify connections to known AI services. These signals establish contact and usage patterns; they may not identify the account, prompt or uploaded document.
Check expenses and procurement. Small subscriptions can reveal services that missed intake. Reconcile them with approved supplier and software records, but remember that free accounts and bundled features may leave no purchase trail.
Inspect identity, endpoint and development records. Review OAuth grants, browser extensions, installed AI tools, local runtimes and model API integrations through the appropriate administrators. Distinguish permission granted from permission actually exercised.
Review enabled AI features in existing SaaS. Ask application owners which features are active and check configuration, release notes, provider notices and data-flow documentation. A product announcement alone does not establish that your tenant uses the feature.
Ask employees about the work. A short, non-punitive survey can identify tools or workflows the technical signals miss. Ask which account they use, what information they provide, and what an approved alternative would need to do.
Reconcile findings into one inventory. Group observations by use case, account and owner. Keep links to evidence and record unresolved questions. Avoid counting the same service once for a domain visit, again for an expense, and again for an OAuth grant.
Microsoft’s shadow AI discovery documentation illustrates the distinction: network-based application discovery is separate from inspection of prompt content and MCP operations. An application inventory does not, by itself, prove what data crossed the connection.
Turn observations into an inventory you can review
A discovery record should make it possible for another reviewer to tell what was observed and what remains unknown. Keep the underlying evidence restricted to the people who need it; collecting more prompt content is not automatically better governance.
For each use, record:
The service, feature or package, plus the account, workspace or identity used.
The business owner and workflow: employee assistance, an internal process or a production service.
The data and permissions involved, labeled as confirmed, reported or unknown.
The evidence source, observation date and collection coverage.
The approval status, next action, responsible reviewer and trigger for reassessment.
Prioritize confirmed sensitive uploads, broad delegated access and agents able to change records or disclose information. A visit to an AI vendor’s website should not receive the same treatment as an agent connected to a customer database.
Measure coverage as well as the number of tools
A falling tool count could mean consolidation. It could also mean activity moved outside the sensors. Track coverage of managed devices, identities, networks and SaaS tenants alongside findings. Review records with no owner, unknown data flows or overdue decisions.
Sample approved uses too. Check whether the account and permissions in use still match the approval, and whether new features or integrations have changed the workflow. A product can remain approved while a particular use falls outside its agreed scope.
Give each finding a next step
Discovery is complete enough to act when the team can describe the use, identify its owner and explain the uncertainty that remains. Our guide to reviewing shadow AI data flows covers the next decision: approve a defined use, impose conditions, move it to a managed account, or stop it.
When an external provider needs review, use the AI vendor assessment questions to request evidence. Internally operated tools need engineering and security owners; a supplier questionnaire cannot answer questions about your own deployment.
Discovery feeds a wider program. The enterprise AI governance guide explains how to connect the inventory to ownership, risk review and ongoing monitoring.
The aim is a record people can use: what is running, what it can reach, what was approved and who will resolve the remaining questions.





