AI councils & champion programs

Most organisations already have steering forums for budget and platform teams for standards. What fails is the middle layer: who prioritises use cases, who coaches business units, and who translates NIST AI RMF risk into practical guardrails.

An AI council aligned to NIST AI RMF Govern sets direction and decides what gets funded. Champions carry patterns into departments, run discovery, and feed the council real adoption signals. Together they beat a single central team trying to own every workflow.

A centre of excellence without governance prioritisation becomes a ticket queue. A council without champions becomes slides without adoption.

Federated business units need champions with local credibility; pair them with a shared playbook vocabulary across business units.

Avoid duplicate forums. A clean split: executive steering approves funding and risk appetite; AI council runs intake, standards exceptions, and cross-initiative dependencies; CoE / platform ships eval harness, AI Gateway, and telemetry.

If the council only hears status updates, it becomes a calendar tax. Every meeting needs one decision: prioritise, pause, or change a guardrail.

Steering wants outcomes; councils want comparability across intake forms; platform wants reference architecture patterns patterns.

Escalation: unresolved disputes go to executive steering within one week per program charter.

Publish a one-page charter: purpose, membership, decision rights, escalation, and what the council will not do (not replace legal or architecture board).

Sponsors sign in week one so disputes reference charter decision rights, not hallway consensus.

Include quorum and minutes storage for ISO 42001 management review evidence.

Review charter annually or when production readiness conversation or regulation changes.

Use a lightweight intake template so every idea arrives comparable: problem owner, users, data sources, read vs write, success metric, and stop rules. Score on value, feasibility, risk, and reuse of platform patterns.

Cap active pilots (often three to five at medium scale) so champions are not spread across fifteen demos. Finish or stop before starting new headline pilots.

Publish prioritisation criteria to the organisation. Opaque scoring breeds politics and shadow projects.

Link intake fields to the pilot scoping guide so approved ideas arrive sprint-ready.

Pick champions for credibility in the business, not only model enthusiasm. They should know the workflow and tolerate saying no per AI security controls.

Protect time in role descriptions; monthly forum with platform and council chair.

Rotate after 12–18 months with handover to program templates.

Executives who assign champions without freeing time should expect shadow Copilot use.

Champions are not unpaid consultants. Their job is discovery, adoption, and feedback loops the council cannot see from the centre.

They socialise approved patterns: RAG (Azure RAG concepts) with citations, HITL (OpenAI safety best practices) before writes, eval rubrics, security baseline. They collect override reasons, empty retrieval cases, and workarounds that signal shadow AI.

They bring one demo or OpenAI evals story to the council per quarter. Stories beat aggregate statistics.

They do not own production on-call unless funded per production readiness conversation.

AI programmes fail when security is a late gate. Reserve standing time for exceptions, incidents, and shadow-AI per AI security controls.

The council does not replace InfoSec sign-off but ensures every pilot uses the AWS AI compliance checklist.

Approve new tools or data classes only with expiry. Permanent exceptions become ISO 42001 audit findings.

Review kill-switch drill status before production readiness conversation scale.

Champions fail when guardrails are vague. Ship an enablement kit: model routes (AI Gateway), data privacy rules, tool templates, eval starter set, security controls.

Provide sandbox with telemetry and Content Safety defaults matching production.

Version the kit like release notes when reference architecture patterns patterns change.

Train champions on half-day pilot scoping workshops using reference architecture patterns.

Predictable cadence beats ad hoc Slack. Calendar invites with agendas reduce programme fatigue per program office norms.

Monthly council meetings focus on decisions. Fortnightly champion community focuses on peer learning. Weekly platform office hours unblock choosing cloud integrations.

Quarterly refresh: risk appetite, production readiness conversation, champion roster.

Publish minutes in 48 hours for ISO 42001 management review traceability.

Calling out failure modes early builds trust with sceptical leaders.

Veto-only councils slow pilots without an enablement kit. Champions as shadow admins rotate keys.

No stop rules let every pilot become permanent beta. Metrics theatre without OpenAI evals sampling.

Ignoring Copilot coexistence wastes council credibility.

Track outcomes the executive steering forum cares about plus signals champions influence via OpenAI evals.

Portfolio metrics: pilots with scale, pivot, or stop documented. Reuse: workflows on AI Gateway and eval patterns.

Risk metrics: incidents, override rate, shadow-AI resolved to security controls paths.

Tie cost to unified model gateway on one dashboard.

Half-day launch with executive sponsor, council members, champion cohort, and platform lead.

Morning: charter, intake template, weights, exceptions. Afternoon: champions, enablement kit, calendar.

End with three pilot scoping charters and security liaison assignments.

Book first council meeting and champion community before attendees leave.