Evidence Grounded AI Control Plane for Enterprise Storage

Enterprise storage teams are beginning to use AI agents to interpret telemetry, investigate incidents, and recommend operational changes. Directly connecting a language model to administrative APIs, however, creates an unsafe coupling between probabilistic reasoning and authoritative infrastructure control. This paper proposes an evidence-grounded AI control plane for distributed storage operations. The architecture separates an AI reasoning plane from a deterministic authority and execution plane. Agents collect metrics, logs, traces, topology, and change history; compare competing hypotheses; and produce a structured action contract. An independent policy gateway verifies identity, scope, preconditions, reversibility, approval, and expected outcome before an allowlisted executor can act. Every operation is idempotent, auditable, and followed by authoritative outcome verification. The framework supports a staged path from read-only diagnosis to bounded remediation while addressing prompt injection, excessive agency, stale telemetry, and cascading actions. It is a reference architecture and does not claim a production deployment or measured operational improvement.

Artificial Intelligence Cloud Computing Software Engineering