Enterprise AI Risk Governance: Four Defenses for Production

Governance is continuous risk control

AI risk is broader than hallucination. Correct text can still cause harm if it exposes restricted knowledge, uses the wrong identity, triggers an unconfirmed action, or cannot be audited. Trusted AI means risks are identified, limited, detected, and recoverable.

Defense 1: data and identity

Define source, purpose, retention, owner, and sensitivity. Apply minimization, masking, approved tools, credential storage, and least privilege. Propagate user identity through retrieval and tools, and invalidate indexes when source access changes.

Defense 2: model input and output

System prompts are not a security boundary. Isolate external content, test prompt injection, filter sensitive input, verify important facts, bind claims to evidence, and allow refusal or escalation. High-impact decisions require meaningful human review, appeal, and accountability.

Defense 3: tool execution

Validate parameters in deterministic services. Enforce permissions, idempotency, rate limits, timeouts, logs, and rollback. Reading public information may be automatic; payments, deletion, external commitments, price changes, and access changes require confirmation and batch limits.

Defense 4: suppliers and operations

Review model APIs, proxies, open-source components, data terms, availability, pricing changes, and exit options. Monitor failures, latency, cost, blocked sensitive requests, tool calls, and complaints. Prepare isolation, credential revocation, fallback, evidence preservation, notification, and recovery.

Set control strength by consequence: low-risk drafts can move quickly; high-impact automated decisions require stronger evaluation, approval, and monitoring. Governance shows teams where they may move fast and where they must slow down.

Based on the local notes “Risk Avoidance,” “Three AI Production Gaps,” and “AI Project Implementation Method.”