Detecting shadow AI requires more than scanning for application names. The same tool can support a harmless public-data task or a sensitive, consequential workflow. A useful discovery process combines technical signals with direct conversations about how work is changing.
Start with a clear purpose
Tell employees that discovery is intended to provide safer tools and better support. Define who can see the findings, how long data is retained, and how disclosed experiments will be handled. People are more likely to share useful information when the process is predictable and non-punitive.
Use six evidence sources
Employee and manager interviews
Ask about tasks, not favorite tools. Which steps are repetitive? Where are people pasting information into a model? Which outputs reach customers, employees, or regulated processes? Interviews reveal browser tools, personal accounts, and informal workflows that technical systems may miss.
Expense and procurement data
Search reimbursements, cards, app catalogs, and vendor requests for AI subscriptions. This shows paid use but will not capture free tools.
Identity and access signals
Review sanctioned application access, single sign-on registrations, OAuth grants, and connected applications. Focus on permissions and data reach rather than raw login counts.
Browser and network signals
Where policy and local law permit, aggregate domain-level usage or managed-browser extension data. Use the minimum detail needed to identify exposure. Do not turn discovery into employee surveillance.
Data-loss and security alerts
Existing security systems may identify uploads, sensitive text, unusual API activity, or unsanctioned connections. Validate alerts with workflow context before assigning severity.
Workflow reviews
Sample real processes in sales, finance, operations, HR, engineering, and customer support. Ask where AI is used to draft, summarize, classify, recommend, decide, or act.
Build a workflow register
For each use case, record the business purpose, owner, tool, account type, input data, connected systems, output audience, degree of autonomy, current human review, and known incidents. Then assign a temporary decision: approved, approved with conditions, needs assessment, or stop pending review.
Avoid treating a domain list as the final inventory. The unit of governance is the use case because risk depends on the work being done.
Prioritize the first response
Address workflows first when they use restricted data, affect rights or material decisions, publish externally, create code or payments, have broad system access, or operate without a clear human owner. Low-risk experiments can move through a faster path.
Track the right measures
Useful measures include disclosed workflows, percentage with accountable owners, time to review, migration to approved environments, repeat policy questions, and high-risk uses closed or controlled. A healthy program may initially show more shadow AI because visibility and trust improve.
Detection is successful when it converts hidden activity into informed decisions and safer alternatives—not when a dashboard reports zero unauthorized tools.