A practical decomposition of the CoE's operating shape, adapted from Automation Anywhere's pillar model.
Pillar 1 — Strategy & prioritization
- Maintain a Value-vs-Risk matrix for incoming requests.
- Prioritize cross-functional, high-complexity workflows (claims processing, supply chain exceptions, end-to-end onboarding) over isolated productivity gains.
- Every approved initiative maps to a specific business KPI (cycle time, FTR rate, cost saved, error rate reduced).
- Distinguish "needs an agent" from "needs a better API" or "needs a deterministic workflow." Not every problem is an AI problem.
Pillar 2 — Embedded governance & agentic guardrails
- Autonomy thresholds are concrete and per-process. ("Agent approves refunds <$500 autonomously; $500–$2,000 routes to manager; >$2,000 is out of scope.")
- Chain-of-thought / reasoning logs are captured for every agent decision, not just final outputs.
- Red-team protocols test models for prompt injection, jailbreak, and unsafe-content emission before production.
- Policies travel inseparably with the work — enforced at runtime, not retrieved from a wiki at audit time.
Pillar 3 — Unified architecture & LLMOps
- The CoE owns the reference architecture: orchestration layer, LLM selection, RAG approach, vector store, observability, identity.
- Agents use least-privilege access to systems of record via scoped, monitored connectors.
- Model drift, embedding drift, and tool-call drift are monitored. Performance degrades silently with data changes; the architecture must detect this.
- The architecture supports multi-model. We do not lock the entire portfolio to one LLM vendor.
Pillar 4 — Enablement via reusable building blocks
- Governed prompt templates with built-in safety instructions, versioned and deprecated like code.
- Agent blueprints / personas (e.g., "Customer Service Triage Agent") that workload teams customize.
- Workflow orchestration patterns for common integrations (legacy CRM, ERP, ticketing, file shares).
- A knowledge hub (internal portal) for templates, runbooks, retro reports, eval datasets.
- Active lifecycle management — versioning and deprecation. A drifting library is worse than no library, because teams route around it.
Pillar 5 — Measurement & continuous feedback
Traditional ROI alone is not enough for agentic AI. Track:
- Decision quality — agent judgment accuracy vs. human benchmarks.
- HITL frequency — how often the agent escalates. Trending to zero is a red flag, not a success.
- Token efficiency — cost-to-value per execution.
- Adoption — unique users, repeat use, time saved.
- Incidents — false outputs, blocked actions, prompt-injection attempts, policy violations.