A customer sends a security questionnaire that asks which AI systems touch their data. The team answers from memory and lists two. A later review of single sign-on logs, expense reports, and browser extensions turns up many more: a meeting summarizer, a coding assistant, an AI feature inside the support desk, and a model a data scientist called from a personal account.
Most organizations have no single list of their AI systems. Each team knows its own tools, nobody owns the whole picture, and a ban on unapproved tools simply pushes use out of sight. Auditors and customers increasingly ask for an inventory, and an incomplete one is worse than a known gap because it looks like you did not look.
RAMP builds the AI register from three intake paths: employee self-reports, imports from your identity provider, and spreadsheet uploads. Each item is triaged into the register as an AI system, a non-AI tool, or a duplicate, and every approved system moves on to a requirements document and two-stage approval. The register becomes the starting list for your risk assessment and your audit package.
What counts as shadow AI
Shadow AI is mostly not malicious. It is people trying to work faster with tools nobody told them how to request.
Three places to look
What to do with what you find
Frequently asked questions
- What is shadow AI?
- Shadow AI is AI that people use for work without the security, legal, or risk teams knowing about it or having approved it. It includes standalone tools, AI features inside approved software, browser extensions, and models called from code.
- Why is shadow AI a risk?
- Unreviewed AI can receive confidential, personal, or customer data, produce output nobody validates, and create obligations under contracts and regulations that the organization does not know it has. It also leaves gaps when a customer or auditor asks for an AI inventory.
- How do you find shadow AI in a company?
- Combine employee self-reports, identity provider application data, and procurement or spreadsheet records, then triage everything into a single AI register with an owner for each system.
Do not start with a ban. Start with an inventory: collect from the three sources, triage into one register, and give every system an owner. You cannot govern, price the risk of, or evidence what you have not listed.