The artifact library and the operating rhythm
Lesson 5 of 6 in The AI Center of Excellence: Organizing for AI Adoption.
A center of excellence that leaves no artifacts behind was a consulting engagement. The AWS AI/ML CoE guidance sorts CoE outputs into two families — guidance (best practices, lessons learned, tutorials) and capabilities (skills, tools, reusable templates) — and the artifact library is where both accumulate. Every template the CoE publishes is a decision made once and reused everywhere: the intake form is the intake policy in executable form; the gate checklist is the release standard; the runbook is the incident procedure. Organizations that skip the library re-litigate every question per project, which is how a five-person center ends up in forty meetings a week.
Thirty-odd artifacts sounds like bureaucracy until you group them: five categories, each answering one question the organization will otherwise answer badly under pressure.
Foundation — who are we and who decides?
The charter (mandate, scope, decision rights, reporting line — the document lesson 2 designed), the operating-model description (centralized, hub-and-spoke, or federated, with the migration triggers written down), the RACI for the AI lifecycle, and the funding model (central budget, chargeback, or hybrid — decided before the first invoice, because funding fights retroactively fought are lost by the CoE). Four documents, drafted in the first month, revisited annually.
Governance — what are the rules and how do we prove we follow them?
The AI policy and acceptable-use policy, the risk-tiering rubric, the registry schema (fields, owners, update triggers), the impact-assessment template per tier, the exception process (time-boxed, signed, logged), and the incident runbook with severity classes and escalation clocks. These are the artifacts a regulator or auditor will ask for first — and the ones that map most directly onto an ISO/IEC 42001 AIMS: an AICoE that maintains this shelf is already operating most of the management-system clauses, which is why the AIMS module calls the CoE its natural operational home.
Delivery — how does a use case move?
The intake form and scoring scorecard (value x feasibility x risk), the gate checklists for each of the four gates, the evaluation criteria per tier, the pattern library (approved reference architectures: RAG on the approved stack, tabular scoring, agent-with-tools — each pre-cleared by the ARB), and the approved-tooling catalog with the request path for adding to it. Delivery artifacts are the enablement half of the mandate made concrete: each one converts a review question into a self-service answer.
Enablement — how do skills leave the center?
The training curriculum by role (executives, builders, reviewers, everyday users — the role-based design that also satisfies AI literacy obligations like EU AI Act Article 4), the champions playbook (selection, training, time commitment, recognition), and the front-door guide — the one-pager that answers 'I have an AI idea, what do I do?' in under a minute. If the front-door guide takes longer to find than a vendor signup page, lesson one’s spiral has already begun.
Measurement — is any of this working?
The KPI tree (adoption, value delivered, risk posture, delivery speed — with the coverage-vs-effectiveness discipline from the operating-model module), the dashboard specification feeding monthly and quarterly reviews, the maturity model the center assesses itself against annually, and the quarterly portfolio review deck template. Microsoft’s responsibility list is explicit that the CoE measures and reports outcomes — adoption rates, compliance levels, project cycle times — to leadership on a regular rhythm.
Artifacts are the CoE’s nouns; cadences are its verbs. Microsoft’s agentic-CoE guidance builds the operating rhythm explicitly — weekly, monthly, quarterly — because a CoE without a rhythm becomes the second failure mode on its list: the ghost, a center with no operating rhythm, no decision authority, and no active membership. Four meetings, each with a distinct altitude, cover the year:
| Cadence | Who is in the room | What gets decided | The artifact it feeds |
|---|---|---|---|
Weekly office hours | CoE on-duty pair + anyone who shows up | Nothing formally — questions answered, ideas pre-triaged, blockers unstuck before they become exceptions | FAQ updates, front-door guide fixes, intake pipeline seeding |
Monthly governance review | CoE governance lead, spoke leads, second-line partner, legal/privacy liaison | Gate decisions in the delegated lane, exception approvals and expiries, incident postmortems, registry hygiene | Registry updates, exception log, calibration notes for spokes |
Quarterly portfolio review | CoE lead, executive sponsor, steering committee, business sponsors | Kill/fund/scale calls across the pipeline, backlog reprioritization, funding-model adjustments | Portfolio deck, KPI tree actuals, next-quarter roadmap |
Annual maturity assessment | CoE core team + an outside pair of eyes (internal audit or an external assessor) | Where the program sits on the maturity model, which operating-model migration triggers have fired, what the center should stop doing | Maturity report to the board, next year’s charter revision |
Tool: AICoE Starter Kit — Stock your own shelf: the AICoE Starter Kit generates all thirty artifacts — charter to maturity model — as editable templates you can download and adapt.
Interactive checkpoint quiz (2 questions) — open this page in a browser to take it.