The idea

AI is now a cross-cutting concern like identity, logging, and security. Put it on the diagram. Give it owners. Give it budget.

Architecture pattern

• Data contracts: schemas, PII policy, lineage, retention.
• Retrieval layer: search/embedding indexes with access control.
• Orchestration: tasks/agents with human-in-the-loop and audit.
• Evaluation: offline tests + production metrics (quality, cost, time).
• Safety: rate limits, guardrails, red-team, incident playbooks.

First 30–60 days

• Pick two “annoying but safe” use cases (e.g., search, summarization).
• Stand up a retrieval service; wire to one system of record.
• Add human review; log prompts/outputs; define drop-in/out path.

Signals / KPIs

• Cycle time saved per worker; first-response time; cost/request.
• Accuracy vs. baseline; percentage auto-approved; rework rate.

Risks & mitigations

• Privacy: enforce contracts; mask at source; least-privilege.
• Drift: periodic evals; shadow tests; versioned prompts/models.