The problem
Shadow AI is not malice, it is convenience. A developer pastes production data into a chat window; a script ships with a hard-coded provider key; a new tool sends prompts to a region you never approved. Each is invisible to security and finance until something goes wrong.
How it works
Give every team and app one endpoint to call, with a virtual key instead of a raw provider key. Tie access to your enterprise identity through SSO or a validated token, so every call carries a real person or service behind it. From then on, the sanctioned path is the path of least resistance, and everything on it is governed.
export OPENAI_BASE_URL=https://gateway.aigw.app/v1
export OPENAI_API_KEY=aigw_your_key
What you get
- One sanctioned, OpenAI-compatible endpoint that teams actually want to use.
- Every call attributed to a person, team, and project, with the model and data categories it touched.
- Policy, data-loss controls, and budgets applied uniformly, so a new team inherits your guardrails on day one.
- A full audit trail of who used which model, so the answer to “what are we running through AI” is a query, not a survey.
- Provider keys held in the gateway, so no app ever ships with one again.
Start with the quickstart and keys.