Auditable AI reporting: from prompt to UQL

Why "auditable" matters

Letting an AI answer questions about your business data is only useful if you can trust — and prove — what it did. A black-box answer that nobody can reproduce is a liability, not an insight. Nexus is built so that every AI-generated answer reduces to queries you can read, re-run, and govern.

Grounded in your real schema

Before it answers anything, an agent discovers the connected system's collections and fields, so it works from your actual schema rather than guessing table or column names. It then composes UQL — the same structured query language a human analyst would use — and executes it against your data.

Review before it runs

Because the intermediate artifact is plain UQL, you can inspect exactly what will execute. The agent can save its work as a parameterized report, which is then version-controlled, RBAC-gated, and re-runnable by name — turning a one-off prompt into a durable, reviewable asset.

Inside your tenant, under your permissions

Generation and execution happen entirely within your Azure subscription, through the MCP server, under the signed-in user's RBAC. The agent can never exceed that user's access and never sees raw tokens.