A water utility wants every department to consume foundation models through shared observability and policy controls instead of each team calling Bedrock differently. Which enterprise pattern best provides that?

A centralized GenAI gateway or abstraction layer lets a water utility give every department shared observability and policy controls instead of each team calling Bedrock with different clients. The gateway standardizes authentication, routing, guardrail hooks, retries, and telemetry so finance, field ops, and customer service consume foundation models consistently. Fragmentation disappears: one place enforces municipal AI policy and one place shows usage, errors, and latency. Letting each department create its own unmanaged Bedrock client with unique retry and logging conventions fails the shared-control requirement and produces uneven reliability and audit gaps. Banning all department apps from using GenAI and forcing staff onto personal consumer accounts fails enterprise governance—shadow AI is worse than a governed gateway. Publishing one global IAM user access key for every utility employee to share fails identity, rotation, and least privilege at once. Exam caveat: a gateway is a control plane, not a free pass to send sensitive SCADA or customer PII without classification rules. Design routing so high-risk prompts hit stronger Guardrails and human review. Avoid making the gateway a single unmanaged bottleneck without HA and quotas. Operational check: Route departmental apps through the gateway only; verify mutual auth and per-department credentials or API keys; confirm centralized logs land in the utility SIEM; enforce Guardrails and rate limits at the gateway; and periodically compare direct Bedrock CloudTrail calls against gateway traffic to catch bypasses.