Should a Team Train One AI Champion or Distribute Practice Across Roles?

The training plan often starts with a staffing question: who owns it? That sounds simple until one person becomes the only source of examples, answers, and approval.

The better question is where practice must happen. The answer changes with workflow variance, risk, learner volume, manager time, and the speed at which mistakes need to reach someone who can fix them.

Sourceable artifact: This worksheet recommends a hybrid for most rollouts, but it also shows when distributed practice wins and when a central champion should coordinate rather than teach alone.

Should a team train one AI champion or distribute practice?

Use a hybrid by default. Give a central champion responsibility for governance, shared standards, reusable examples, sequencing, and cross-team feedback. Give role-near people responsibility for practising the work, translating examples, and surfacing corrections.

That split follows the distinction in OpenAI's AI Champion role guidance: Leaders coordinate strategy, governance, rollout, and enablement, while Activators work inside teams and adapt AI to specific workflows. OpenAI's leadership guide also recommends both role-specific training and an AI champions network, not one in place of the other (OpenAI's leadership guide).

This is the allocation decision inside the broader AI learning capability guide: decide who owns shared standards, then place practice and correction as close to the work as the risk allows.

One champion can carry most enablement when the workflow is shared, local variation is low, risk is high, learner volume is manageable, and the champion is not the only person who can review or explain the work. In that case, central coordination is useful. Central-only learning is still a weak design if nobody else practises.

Score the operating context before choosing the model

Score the fit of each operating model from 0 to 5 across six criteria. Use 0 when the model conflicts with the context, 3 when it can work with compensating controls, and 5 when it naturally matches the context. Give each criterion equal weight. Your total is out of 30.

Criterion	Ask before you score	A high-fit model lets you...
Workflow proximity	How much does the work vary by role, queue, customer, or judgment call?	Practise on the task as it really occurs.
Need for local translation	How much do policy, vocabulary, data, or tools change by role?	Adapt a verified example without waiting for the center.
Governance risk	What is the consequence of leakage, wrong output, or unreviewed use?	Keep permissions, review, and escalation consistently owned.
Learner volume	How many people, functions, locations, or languages are involved?	Reach the audience without creating a support queue.
Manager capacity	How much protected time exists for coaching and review?	Avoid assuming time that managers do not have.
Feedback speed	How quickly must a bad example or workflow issue reach an owner?	Return questions and failures to someone who can act.

These criteria are not arbitrary training preferences. OpenAI describes workflow-near Activators, the GOV.UK employer evidence base emphasizes practical and contextualized learning, and Microsoft's Copilot rollout combined central change management with departmental teams and role-specific content. The GOV.UK evidence base draws on 23 workshops, 10 case studies, and a 536-response survey (GOV.UK employer guide).

Use this decision rule:

1. Add the six scores for each option.
2. Treat 24 or more as a candidate recommendation, 19 to 23 as a sequencing or pilot signal, and 18 or less as a reason to redesign the model.
3. Let a veto override the total.
4. Re-score after the first practice cycle. A change in risk, learner volume, manager time, or feedback speed can change the choice.

The worksheet is a transparent starting rule, not an empirical winner. If you have real numbers, use them. If you do not, mark an uncertain row as 3 and record what would change the score.

Three worked choices

The scores below are illustrative. They show how the same three options behave in different operating contexts. They are not client results.

Low-variance, high-risk work

Imagine one controlled workflow, a small learner group, sensitive data, limited manager time, and a strong need for consistent review.

Criterion	One central champion	Distributed practice	Hybrid
Workflow proximity	4	3	5
Local translation	4	2	4
Governance risk	5	2	5
Learner volume	4	2	4
Manager capacity	5	2	4
Feedback speed	2	4	5
Total / 30	24	15	27

Recommendation: choose the hybrid. The central champion owns the approved workflow, data boundary, review rule, and reusable example. Each learner still practises the same workflow with a named supervisor or reviewer.

Veto: if the champion is also the only approver for every live case, stop rollout. Add an independent review owner before training becomes operational use.

High-variance, role-specific work

Imagine several functions using different systems, vocabulary, customer situations, and quality judgments. Local managers have some protected practice time.

Criterion	One central champion	Distributed practice	Hybrid
Workflow proximity	1	5	4
Local translation	1	5	4
Governance risk	3	2	4
Learner volume	2	3	3
Manager capacity	3	4	3
Feedback speed	2	5	5
Total / 30	12	24	23

Recommendation: distribute practice across roles, with a light central champion for the shared baseline, governance questions, and reusable patterns. Local owners choose examples and collect corrections.

Veto: if no local owner can protect practice time or explain what a good output means, do not decentralize yet. Run a bounded common workflow and create local ownership before expanding.

Scaled multi-function rollout

Imagine many functions and locations, mixed learner experience, shared governance requirements, and a need to reuse materials without erasing local differences.

Criterion	One central champion	Distributed practice	Hybrid
Workflow proximity	2	5	5
Local translation	2	5	5
Governance risk	5	2	5
Learner volume	5	3	5
Manager capacity	4	2	4
Feedback speed	2	5	5
Total / 30	20	22	29

Recommendation: use a hybrid network. The center owns rollout sequencing, minimum standards, governance, a shared resource hub, and cross-function feedback. Functional Activators own role-specific examples, practice sessions, and the first signal that a workflow does not fit.

Veto: if the central team cannot name functional or regional Activators, do not call the rollout hybrid. Start with a smaller pilot or fund local owners first.

Microsoft's own account of its Copilot rollout illustrates this coordination layer. It reports central change-management work alongside departmental teams, a Champions Team, role-specific communities, and local-language training. The case also describes about 30,000 employees attending instructor-led courses, which is a scale problem a single trainer could not solve alone (Microsoft's Applying AI at scale).

What does each model cost?

The cheapest model is not the one with the fewest trainers. Estimate three loads before buying training:

support load = learners x expected help requests x minutes per request
translation load = role or policy variants x minutes to adapt one verified example
governance load = workflows x review points x review time

Central-only reduces duplicated content and makes governance easier to coordinate. It also concentrates support, translation, and institutional memory in one person. A central champion can become a queue.

Distributed practice spends more manager and role-owner time. It can also duplicate examples or weaken common standards. Its benefit is speed close to the work. People do not need to wait for a central trainer to explain the vocabulary, exception, or approval habit that matters in their role.

Hybrid pays for a central coordination layer plus local practice. The initial cost is higher because both layers need time. The trade is that governance and reusable assets have an owner while practice and feedback remain close to the workflow.

The GOV.UK employer guide explicitly treats protected time, local champions, peer support, governance, expandable training, and sustainability as design concerns. That is a useful cost model: count the time needed to make practice reachable and safe, not only the cost of the course.

Why informal practice changes the decision

Formal training is not the only route by which people learn an AI tool. In the original arXiv study of enterprise AI learning, none of the 10 interviewed M365 Copilot users reported formal training as their primary learning method. Eight relied on trial and error, and six exchanged tips with colleagues. The study is exploratory and small. It does not prove that informal learning is better, but it does show why a rollout that ignores peer practice can miss how value is actually discovered.

OpenAI's Champion guidance makes a similar operational distinction: Leaders help create direction and momentum, while Activators turn that momentum into everyday practice. OpenAI's leadership guide recommends a champions network, role-specific training, routine experimentation, a shared knowledge hub, and active internal communities (OpenAI's role guidance, OpenAI's leadership guide).

F-orange firsthand observation: In the ChatGPT workshop I led at Orange, we began with participants' existing work. This is one bounded teaching observation, not a measured adoption result. It is why workflow proximity is a first-class row in the worksheet rather than a soft preference.

What usually fails?

Four failure modes show up in the allocation itself:

1. The champion becomes a help desk. Set office hours, a response route, reusable answers, and a second layer of local owners.
2. A central demo becomes a local policy. Separate the reusable example from each role's data, permission, and review boundary.
3. Distributed practice fragments the standard. Publish a short baseline, approved examples, a review rule, and an escalation path before local experimentation.
4. Attendance is mistaken for capability. Require a real work artifact, a verification step, and a new-task transfer check. The AI learning capability guide explains that distinction in detail, while this guide to making AI training stick in a small team covers the follow-through loop after the session.

Do not choose distributed practice when local owners have no protected time, shared baseline, or escalation path. Do not choose central-only when one champion would be the only person able to translate, troubleshoot, or approve the work. Do not choose hybrid as a label for central training plus optional local activity. Name the local practice owner.

The buying decision

Choose one central champion only when the workflow is shared, risk is high, local variation is low, learner volume is manageable, and the champion has capacity to coordinate without becoming the sole practitioner. Choose distributed practice when role context and feedback speed dominate, but fund the baseline and governance needed to keep local work safe. Choose hybrid for most multi-function rollouts.

Before comparing training providers, fill the worksheet with one real workflow, three representative learner roles, and the proposed review boundary. Marius Manolachi's AI learning and consulting work can help when the team needs to turn that evidence into a practical learning plan. The decision tool above should still let you make the first choice without hiring anyone.