People and roles: staffing the center
Lesson 3 of 6 in The AI Center of Excellence: Organizing for AI Adoption.
Both vendor playbooks start staffing in the same place, and it is not a job posting. Microsoft’s five-step build sequence puts executive sponsorship first — budget, authority, and credibility, because without it the CoE 'can’t enforce standards or drive organizational change' — and step two is a single accountable leader, one named person who owns AI strategy implementation and can influence stakeholders. Only then does the multidisciplinary team assemble. AWS’s AI/ML CoE guidance sketches the same 'dream team' shape: product strategists, AI researchers, data scientists and ML engineers, domain experts from the business units, operations people for the MLOps pipelines, and project managers — flanked by an executive steering committee and, for responsible-AI questions, an ethics board.
Translated into the roles you will actually hire or borrow, a working AICoE core needs seven functions covered — functions, not necessarily seven heads; in a small organization one person wears three of these hats:
| Role | Owns | Fails without them |
|---|---|---|
CoE lead | The charter, the roadmap, the budget, the C-suite relationship — the single throat to choke for the dual mandate | No one arbitrates enable-vs-govern trade-offs; the center drifts toward whichever half its loudest member prefers |
AI governance lead | The risk-tiering rubric, the gate criteria, the registry’s integrity, liaison to the governance board and second line | Reviews become vibes; the same use case gets different answers on different days |
Platform / ML engineers | The shared delivery platform: approved tooling, pipelines, guardrails as code, monitoring plumbing | Every team rebuilds the plumbing; standards exist only as documents nobody can execute |
Data scientists | Use-case feasibility calls, evaluation design, model quality — the craft the spokes learn from | The center cannot judge technical claims and rubber-stamps whatever the vendor deck says |
Legal / privacy liaison | Regulatory triage at intake, contract and data-use review, the bridge to DPO and counsel | Legal review happens at the end, where it is a launch-blocker instead of a design input |
Enablement lead | Training curriculum, office hours, the champions program, internal communications | Skills never leave the center; the CoE scales linearly with its own headcount forever |
Domain champions | Local use-case scouting, first-line triage, translating standards into unit dialect | The center learns about business problems from intake forms instead of from the business |
The core team is only the apex of the structure. What makes a CoE scale is the two layers beneath it: an extended team of part-time specialists who keep their day jobs, and a champions network in the business units. AWS’s CCoE guidance describes exactly this evolution — start with 'a handful of early adopters and cloud champions', then grow champions for both business functions (change management, governance, training, procurement) and technical ones. The champions layer is the difference between a center that serves the organization and one that merely polices it: champions scout use cases before they become shadow AI, answer the easy questions locally, and give the center eyes it could never staff centrally.
The staffing pyramid: small core, wide reach
- Core team — Dedicated — typically 3 to 8 people
Full-time on the CoE: the lead, governance lead, platform engineers, data scientists, enablement lead. Owns the charter, platform, standards, and registry. Small on purpose — every permanent hire here should make the layers below more capable, not more dependent.
- Extended team — Fractional — borrowed hours, formal commitments
Named individuals in legal, privacy, security, procurement, HR, and the data office who owe the CoE defined hours: intake reviews, contract clauses, security assessments. Formalize the arrangement — a name and a percentage in the charter — or these hours evaporate the first busy quarter.
- Champions network — Embedded — one or two per business unit
Volunteers-then-trained practitioners inside each unit: they scout use cases, run first-pass triage, teach the basics, and carry unit context back to the hub. This layer is where hub-and-spoke spokes come from — today’s champion is next year’s spoke lead.
Where do these people come from? Three sourcing channels, in deliberate proportion. Hire the rare skills you will need permanently and cannot grow fast enough — usually the platform engineers and the governance lead. Borrow the specialists whose expertise you need fractionally — legal, security, privacy — because embedding a full-time lawyer in a five-person CoE is neither affordable nor necessary. Train everything else, because training is the only channel that scales: a common pattern is to run the champions program as the deliberate seedbed for future spoke leads, with certifications, hackathons, and mentoring as the curriculum (both vendors’ playbooks lean hard on this — workshops, certifications, hackathons, and mentoring in AWS’s framing; skills assessments, learning pathways, and hands-on experimentation in Microsoft’s).
One warning from the CAF-AI applies to every channel: align the CoE’s incentives 'with the strategy, the business, and the customers'. A center rewarded for papers published or reviews completed will optimize for exactly that. Reward shipped, governed business value — the metric both halves of the mandate share.
Interactive checkpoint quiz (2 questions) — open this page in a browser to take it.