Shadow AI: how to discover and govern AI you don't know about

AI is easy to adopt. Anyone can sign up for an AI assistant, paste a model API key into a script, run an open-weight model on a laptop, or enable an AI feature in a SaaS tool. Most of this happens with good intentions and without telling anyone.

Common forms of shadow AI include:

• Personal accounts on public AI assistants used with company data.
• Model API keys created outside central procurement.
• Open-weight models running locally, for example through Ollama.
• Coding assistants and agents in developer environments.
• AI features switched on inside software approved for another purpose.
• AI agents connected to internal systems without a security review.

Shadow AI is not just a policy violation; it is a control gap. If you can't see an AI system, you can't assess it, monitor it or produce evidence about it.

• Data exposure: personal data, source code or confidential documents may be sent to external models.
• Unassessed decisions: AI outputs may influence decisions about customers or employees without any risk assessment.
• Compliance gaps: your AI inventory, risk register and audit evidence describe only the approved AI, not the real environment.
• Agent risk: unreviewed AI agents may hold access to production systems.

No single data source finds everything. Effective discovery combines several signals and compares them against the approved AI inventory:

• A finance analyst uses a personal Gemini API key to summarize customer files. Discovery flags it; the team moves to an approved, enterprise-licensed model.
• A data scientist runs a local Llama model to process HR data. It is registered, assessed, and approved with restrictions on the data it may use.
• A support team connects an AI agent to the CRM. It is inventoried, its permissions are reduced, and external emails are routed for human approval.

• Disclosure of personal data, trade secrets or source code.
• Unreviewed AI influencing decisions about people.
• Audit findings because the AI inventory is incomplete.
• Inconsistent answers to customer AI questionnaires.
• Duplicate spending on overlapping AI tools.

• NIST AI RMF: the Govern function includes having mechanisms to inventory AI systems, resourced according to organizational risk priorities.
• ISO/IEC 42001: an AI management system has a defined scope; AI that is in use but outside that inventory is effectively ungoverned.
• EU AI Act: obligations depend on an organization's role and on each AI system's risk category, which you can only determine for systems you have identified.
• OWASP Top 10 for LLM Applications: sensitive information disclosure is listed among the top risks of LLM use.

1. Publish an AI use policy with a list of approved tools and a fast way to request new ones.
2. Discover continuously across the sources above, not once a year.
3. Triage each finding: register it, assign an owner, record its purpose and the data it touches.
4. Assess risk and decide: approve, approve with conditions, replace, or restrict.
5. Monitor approved AI so it stays within its conditions.
6. Educate teams on why approved alternatives exist; AI literacy reduces repeat shadow AI.

ReguLattice starts from one principle: you can't govern AI you don't know exists. It discovers AI across model provider APIs, local runtimes such as Ollama, code repositories and developer tools such as Claude Code and Cursor, and compares what it finds with the registered AI inventory.

Unregistered AI is flagged as shadow AI, assigned an owner and assessed. From there, the same platform applies policies, monitors usage at runtime and keeps evidence mapped to ISO/IEC 42001, the NIST AI RMF and the EU AI Act. See how an AI governance platform works.