Shadow AI: How to Discover and Govern Unapproved AI in Your Organisation
Shadow AI is the use of AI tools, models, agents or AI features inside an organisation without the knowledge or approval of IT, security or compliance teams. It ranges from employees pasting data into public chatbots to developers running local models and connecting agents to internal systems.
Key takeaways
- Shadow AI is mostly well-intentioned: people adopt AI because it makes them faster.
- The main risks are data leakage, non-compliance (GDPR, EU AI Act, contracts) and unmanaged agents with real access to systems.
- Discovery combines network, endpoint, identity, code and SaaS signals — no single source sees everything.
- Governance works best with an approved catalogue and safe defaults, not with blanket bans.
Forms of shadow AI
- Public AI assistants used with company data through personal accounts.
- AI features inside SaaS tools switched on by default or by individual teams.
- Browser extensions that read page content to summarise or rewrite it.
- Local models run on workstations or servers by technical staff.
- Agents and MCP servers connected to mailboxes, drives, code repositories or databases.
- API keys for AI providers embedded in scripts, notebooks and applications.
Why it matters
| Risk | Example |
|---|---|
| Data leakage | Source code, customer data or contracts sent to an external service with unknown retention. |
| Compliance | Personal data processed without a legal basis; AI used in a high-risk context without the controls the EU AI Act requires. |
| Security | An unreviewed agent with broad permissions, exposed to prompt injection; leaked API keys. |
| Cost and continuity | Untracked subscriptions and usage; business processes depending on tools nobody owns. |
How to discover shadow AI
| Source | What to look for |
|---|---|
| Proxy, DNS and firewall logs | Traffic to AI provider domains and APIs, volumes per user or team. |
| Identity provider | OAuth consents and sign-ins to AI applications using corporate accounts. |
| Endpoints | Installed AI desktop apps, browser extensions, local model runtimes listening on local ports. |
| Code and configuration | AI SDK imports, provider API keys in repositories, agent and MCP configuration files. |
| SaaS administration | AI features enabled in collaboration, CRM and productivity suites. |
| Finance | Expense claims and card payments to AI vendors. |
Correlate these signals into a single inventory: tool, users, data involved, owner and risk level.
From discovery to governance
- Publish an acceptable-use policy that says clearly what data can go where.
- Offer approved alternatives: enterprise accounts with data-protection terms, or internal models.
- Classify each discovered use by data sensitivity and regulatory impact.
- Apply controls: data-loss prevention on prompts, guardrails, least privilege for agents and MCP servers.
- Monitor continuously: new tools appear every week, so discovery must be ongoing.
- Train staff — AI literacy is also an EU AI Act expectation.
Frequently asked questions
What is shadow AI?
The use of AI tools, models, agents or AI features in an organisation without the approval or knowledge of IT, security or compliance.
Should companies block public AI chatbots?
Blanket bans tend to push usage onto personal devices and make it invisible. A better approach is to provide approved tools, define what data is allowed, and monitor usage.
How can I detect AI agents and MCP servers in my company?
Combine endpoint inventory (agent apps and configuration files), code scanning (SDKs and API keys), identity logs (OAuth grants) and network logs (traffic to AI APIs), then consolidate the results into one inventory.
Is shadow AI an EU AI Act issue?
It can be. An organisation cannot classify, document or oversee AI systems it does not know about, and AI literacy obligations apply to staff using AI.
Related guides
- AI Agent & MCP Security — Threats to agents and MCP servers, with a practical checklist.
- EU AI Act & AI Governance — Timeline after the Digital Omnibus and a governance checklist.
- AI Observability — What to monitor in LLM apps and agents, with OpenTelemetry.