AI Champions Program vs Whole-Team Training: Which Works?

An AI champions program works best as the first phase of an AI rollout, not the whole plan. For most firms whose staff already hold AI licenses, the stronger sequence is a bounded champion cascade: one champion per team spends about eight weeks building working assets on real recurring tasks and teaching peers, then everyone gets role-based training. Train the whole team first when a professional duty covers everyone, the firm has fewer than about 25 people, or a firm-wide tool switch lands on one date. These durations and thresholds are AIDOLS planning estimates.

Key figures (updated 23 October 2026)

• Only 36% of respondents to BCG's 2026 survey felt sufficiently trained, unchanged from 2025 (BCG, June 2026).
• Regular AI use was 18% among employees with no training and 82% among those with five to ten hours (BCG, 2025).
• In Danish AI-exposed jobs, 93% had used AI chatbots where employers combined encouragement, enterprise tools and training, against about 41% where employers took no active steps (Brookings, October 2026).
• In AI-adopting organizations, employees who strongly agree their manager actively supports their team's AI use are 1.7 times as likely to use it a few times a week or more (Gallup, May 2026).
• EU AI Act Article 4 was rewritten by Regulation (EU) 2026/1744, in force since 27 July 2026; national supervision and enforcement apply from 2 August 2026 (EUR-Lex; European Commission).
• In a review of health-care champions, five of a subset of seven studies linked champions to higher uptake; causation is unproven (Santos et al., 2022).

Cite as: AIDOLS Research Team, AI Champions Program vs Whole-Team Training: Which Works?, aidolsgroup.com, 23 October 2026.

This guide covers definitions, evidence, a decision matrix, EU AI Act Article 4, charters, role-based training, measurement, effort, a 90-day plan and templates.

An AI champions program trains a small group first, usually one person per team, who then help colleagues use AI on their own work. Whole-team training puts everyone through a baseline, ideally by role, within one window.

• AI champion network: practitioners who experiment with the firm's AI tools and help colleagues apply them; Microsoft's adoption guidance adds that champions "don't replace IT, compliance, training, or Copilot adoption program teams."
• Whole-team (all-staff) training: everyone completes a baseline, or a role version, in the same window.
• Train-the-trainer: a few people learn to deliver a course to everyone else.
• Champion cascade: the hybrid AIDOLS recommends: champions prove two or three working assets on real recurring tasks and teach peers in short clinics, then everyone gets role-based training.

AI champion network vs AI center of excellence

An AI center of excellence is a central, cross-functional team that sets standards and builds shared platforms; champions sit inside teams and work on the team's own tasks. Microsoft's charter guidance lists what sits outside the champion role: "IT support, licensing, security and compliance approvals, formal training, adoption strategy, or business outcomes." In a 50- or 500-person firm, the "center" can be one program owner plus named owners for licenses, data, policy and outages. Once a firm deploys its own AI systems, end-user training and champions sit in the Scale phase of our AI adoption framework.

Three findings hold across the surveys: regular use is higher with more training and with in-person, coached formats; most staff still lack training and teach themselves; and use is far higher where managers and leaders visibly support it. None of it shows that training alone raises output.

Hours and format: the five-hour line

BCG's 2025 survey of more than 10,600 workers in 11 countries and regions found regular use (daily or several days a week) "sharply higher for employees that receive at least five hours of training and have access to in-person training and coaching" (BCG, 2025). Regular users made up 18% of those with no training, 63% with one to five hours, 82% with five to ten and 89% with more than ten. In-person training went with 79% regular use against 67% without; access to a coach, 84% against 70%.

Most staff are still untrained, so they teach themselves

Only 36% of the 11,749 people in BCG's 2026 survey felt sufficiently trained, unchanged from 2025 (BCG, 2026). Microsoft and LinkedIn's 2024 Work Trend Index found that only 39% of people using AI at work had received AI training from their company, and that 78% of AI users brought their own AI tools to work (Microsoft and LinkedIn, 2024). In BCG's 2025 data, 54% said they would use AI tools even without company authorization. Untrained use is already happening; the question is whether it happens under your ground rules.

Managers and leaders move usage

Employees in AI-adopting organizations who strongly agree their manager actively supports their team's AI use are 1.7 times as likely to use AI a few times a week or more, and Gallup advises organizations to "start by equipping their managers" (Gallup, May 2026). In professional services, a leading sector, Gallup finds only 22% of US employees use AI daily. Among BCG's frontline respondents in 2025, 82% of those with clear leadership support were regular users, against 41% without it, yet only 25% had that support.

Danish data show the same pattern: about 41% of workers in 11 AI-exposed occupations had used AI chatbots even where employers took no active steps, against 93% where employers combined encouragement, enterprise tools and training (Humlum and Vestergaard, Brookings, 2026).

What these numbers do not prove

Most of these figures are correlational, and the time savings are self-reported: 42% of BCG's frontline regular users say they save a full day or more a week. The Danish study found precise null effects of AI chatbots on earnings and hours, ruling out average effects larger than 2%. That is a labor-market result, not your firm's output, but it warns against equating use with value. BCG's own advice to leaders is to "measure value, not adoption" (BCG, 2026).

Three public evaluations of Microsoft 365 Copilot agree: training mattered, multi-format and self-led training did better on confidence and satisfaction, colleagues and managers influenced use, and self-reported time savings did not prove productivity.

For your firm: train in several formats on your own tasks, give support a named owner, train managers as well as users, and never report self-reported minutes as productivity.

It can, but the evidence is thinner than vendor pages suggest. Health-care research links champions and credible peers to better uptake without proving cause, and none of the research reviewed for this guide measures AI champions' effect on output in professional services. Treat champions as a delivery mechanism, not a strategy.

• Implementation research. A 2022 systematic review of 35 quantitative studies of health-care champions found that, in a subset of seven, five reported an association with increased organizational use of best practices, programs or technologies. The authors call for research on causation, note that champions may carry opportunity costs, and advise deploying them with an evaluation plan (Santos et al., 2022).
• Opinion leaders. A 2019 Cochrane review of 24 studies found that local opinion leaders, colleagues seen as credible and trustworthy, probably improve health professionals' compliance with evidence-based practice by a median 10.8 percentage points (moderate-certainty evidence); it could not say whether the way they were chosen matters (Flodgren et al., 2019).
• Train-the-trainer. A 2021 meta-analysis of 16 studies found that train-the-trainer programs may improve trainers' own knowledge, but "No studies measured trainers' ability to deliver the training program" (Poitras et al., 2021). That is why the cascade below adds a transfer check at 30 days.

Start with champions when licenses are in place and use is uneven; start with the whole team when a duty, a deadline or a tool switch applies to everyone at once. When both are true, run a short whole-team baseline in week one while champions build in parallel. The sizes, ratios and durations below are AIDOLS planning estimates, including the core ratio: one champion per team of 8–15 people (about 7–12% of staff), never fewer than two per firm.

Tool maturity is one part of AI maturity: a firm whose people have held licenses for months without changing a single work product is not early, it is stuck.

For US law firms, ABA Formal Opinion 512 says supervisory duties include making sure subordinate lawyers and nonlawyers are trained in the ethical and practical use of the generative AI tools they use (ABA, 2024). The opinion is advisory; Canadian law societies and the SRA set their own expectations, covered in our guide to AI training for law firms.

Not as a mandatory course for every person. Since 27 July 2026, Article 4 requires providers and deployers to "take measures to support the development of AI literacy" of their staff and others using AI on their behalf, taking their knowledge and the context of use into account (Regulation (EU) 2026/1744). It requires measures, but no guaranteed level for any individual. This is general information, not legal advice; check your regulator's or law society's current guidance.

What changed on 27 July 2026

Regulation (EU) 2026/1744, the Digital Omnibus on AI, was published in the Official Journal on 24 July 2026 and entered into force on 27 July, replacing the 2024 wording many pages still quote.

The European Commission's AI literacy Q&A, updated 20 October 2026, says national market surveillance authorities supervise and enforce the rules from 2 August 2026. Penalties are set nationally, and Article 4 is not among the obligations listed in Article 99(4) of the AI Act, so headline AI Act fines should not be read across to AI literacy.

What the Commission says you must and need not do

The Commission's Q&A settles six practical points:

• No certificate is needed: "Organisations can keep an internal record of trainings and/or other guiding initiatives."
• There is no obligation "to measure the knowledge of AI of employees."
• No mandatory trainings are imposed, and "different levels of training or learning approaches could be appropriate."
• Asking staff to read a system's instructions for use "might be ineffective," and relying on them "is therefore not sufficient."
• Staff using ChatGPT to write advertisements or translate should be told about specific risks, "for example hallucination."
• At minimum, ensure a general understanding of AI, and weigh your role (provider or deployer) and the risk of your systems.

Its repository lists more than 40 AI literacy initiatives, yet replicating them "does not automatically grant presumption of compliance with Article 4." A course, ours included, can be one recorded measure; it does not make a firm compliant. See EU AI Act compliance for the wider obligations.

Does it apply to firms in Canada, the US or the UK?

It can. The Act reaches providers and deployers outside the EU "where the output produced by the AI system is used in the Union" (Article 2(1)(c)), and the Commission confirms this holds for Article 4. A firm whose staff use licensed assistants at work is a deployer of them (Article 3(4)). A Toronto, New York or London firm that serves EU clients with AI-assisted work, or operates in the EU, should assume Article 4 may apply.

High-risk uses carry a stricter duty

From 2 December 2027 under the Digital Omnibus, deployers of Annex III high-risk systems, such as recruitment or creditworthiness tools, must assign human oversight to people "who have the necessary competence, training and authority" (Article 26(2)). Most professional-services use of general assistants is not high-risk.

Choose champions on evidence of repeated use and on whether colleagues already copy them, not on enthusiasm or seniority. Then bound the role with a written charter: one named team, hours agreed with their manager, an end date, and a refusal list.

Choose on evidence, not enthusiasm

Rank candidates first on repeated use of something they built, then on whether they cover a team with nobody else. Check that peers already borrow their work (the quiet person whose spreadsheet everyone uses) and that they control the step they will change; without the template or the sign-off, a champion produces a good demo and no change.

Write a bounded charter

The AIDOLS charter, taught in the champion cascade lesson and given in full in the templates below, sets an eight-week term and a refusal list that sends licensing, data, policy and outage questions to named owners the same day (AIDOLS planning estimates). Microsoft's guidance likewise recommends a one-page charter that states the expected time commitment.

Recognition, and the warning signs of burnout

Recognize the work where it counts: a line in the champion's performance review, credit when the team's measure moves, and agreed hours protected from billable or utilization targets. A badge is not recognition if the work stays invisible to their manager.

Watch for helpdesk drift: the champion channel becomes a queue of access and "is this allowed?" questions, replies slow and then stop, the champion's own asset goes stale. The health-care review flagged opportunity costs; in a fee-earning firm they appear as lost client hours. Route out-of-scope questions within the day, and end the term on the date written down.

Champions spend weeks 2 to 5 building and proving two or three working assets on real recurring tasks, then weeks 5 to 10 teaching their team in short clinics, while a program owner runs weekly office hours (AIDOLS planning estimates).

What champions build first

• A reusable instruction for one recurring task: the job, inputs, output shape and standard.
• A shared setup (a project, Gem or agent) that holds the instruction and non-confidential reference material.
• A timed baseline of the task today, and the record it comes from.
• A draft one-page ground rule for the team's data.

How champions teach peers

Each clinic runs on the team's live work: colleagues bring a real input, build or adapt the asset, and save it where they will find it on Monday. The AIDOLS session design behind training that survives Friday puts 70 of 90 minutes on keyboards, and a 30-minute clinic should keep a similar ratio.

How a champion shares a working setup, by tool

Anyone with access can see a shared asset's instructions and files, as Google and Anthropic document, so keep confidential material out and check what your administrators allow. Vendors also retire and rename these containers, as our guide to model-agnostic AI training dates, so keep a plain-text copy of every shared instruction.

Every role needs the same core (what the tools do and where they fail, what data may go in, how to check output), then a session on that role's own recurring work. The Commission accepts different levels of training for different groups, with a general understanding of AI as the minimum.

These four audiences are our recommendation; the AIDOLS course itself has three role tracks (Leadership; L&D and People; Physical and Industrial Operations), and its lesson on choosing a track by the decision you own helps place people.

• Managers are the biggest lever (Gallup's 1.7× finding) and must learn to spot AI-assisted work: 61% of managers in Australia's pulse survey could not confidently identify Copilot outputs (Digital Transformation Agency).
• Frontline professionals need verification sized to the cost of being wrong: 52% of BCG's 2026 respondents say AI has increased their time reviewing and correcting its output (BCG, 2026).
• Support staff need four checks for every paste: the class of the data, the account it goes to, who can retrieve it later, and how long it is kept.

Count a change in the work, not logins, seat counts, attendance or quiz scores. Pick one work product per team, name the record where it is already counted, and check it at 2 weeks, 30 days and 90 days: the AIDOLS method in measuring adoption honestly.

Activity metrics rise both when people are productive and when they are confused and retrying, so they settle nothing. Quiz scores measure recall, which Article 4 does not require you to test. Self-reported time saved is weakest of all: the UK business department "did not find evidence that time savings have led to improved productivity."

Write each measure as one sentence: "By [date], [role] will produce [work product] in [time], counted from [record]." A hypothetical example: by 15 January, the two coordinators on the leasing team will produce the weekly portfolio report in under 40 minutes, counted from document-system timestamps. If counting it means asking people how they feel, it is not a measure.

A champion cascade asks about 14–16 hours of each champion in weeks 2–10 (the 228-minute course path, 2–4 hours building, then two hours every two weeks), 2–4 hours a week of a program owner for 13 weeks, and the role hours in the table above for everyone else, plus three 30-minute clinics for frontline staff (AIDOLS planning estimates). With clinics counted, frontline staff end up near or past the five-hour line that BCG's 2025 survey associates with regular use.

Whole-team-first takes less calendar time, about 2–6 weeks (AIDOLS planning estimate), but no less time per person.

Waiting also costs money, because unused seats are still paid for. For per-seat prices as of October 2026, see Microsoft 365 Copilot, ChatGPT Enterprise and Claude compared; for a Copilot plan that reads Microsoft's usage reports and prices idle seats, see our Microsoft Copilot adoption plan.

Seven steps take a firm from licensed-but-uneven use to a measured change in the work in 13 weeks (AIDOLS planning estimates).

1. Weeks 0–1: name the task and the record. Each team names one recurring task, times it and picks the record the count will come from. The lesson on naming the one task worth handing off gives four tests: recurring, bounded, low-stakes if wrong, and yours. If nobody can describe the workflow end to end, the free workflow map returns a one-page map and a written verdict within three business days.
2. Week 1: set the ground rule and the owners. Publish a one-page AI ground rule for each team and name owners for licenses, data, policy and outages. Under a professional duty or EU exposure, add a short whole-team literacy session this week.
3. Week 2: choose champions and sign charters. Agree the time with each champion's manager in writing and set the end date.
4. Weeks 2–5: build and prove. Champions build two or three assets on live work; at the 2-week check, keep only what was reused.
5. Weeks 5–10: teach. Run clinics every two weeks, weekly office hours and Friday notes; at 30 days, count the people outside the champion group who produce the work.
6. Weeks 8–12: train everyone by role, so each person leaves with a working asset saved where they will find it.
7. Week 13: measure, renew, repeat. Read the 90-day change, renew charters only for champions who ask, pick the second wave, and file a dated record of every session as evidence of your Article 4 measures.

Copy these and fill the blanks; they work whether or not you ever talk to AIDOLS.

Champion charter (one page)

Nomination criteria

Nominate people for whom at least five of these six are true, including the first and the last:

• Have they reused something they built, at least twice, unprompted?
• Do they cover a team with no other candidate?
• Do colleagues already borrow their templates or methods?
• Do they own, or can they change, the step the team's work depends on?
• Is their team stable for the next eight weeks?
• Has their manager agreed the time in writing?

Weekly cadence

A simple AI literacy record

The Commission accepts an internal record in place of certificates. Keep one row per measure.

When the risk or the deadline covers everyone at once, when the firm is too small for a champion layer, or when nobody owns the questions champions will be asked. And neither approach works yet without recurring tasks to build on, an agreed reason for doing this, and licenses, data access and a ground rule in place.

Cases for whole-team-first

• A duty covers everyone who uses the tools, such as the supervisory training expectation in ABA Opinion 512.
• The firm has fewer than about 25 people (AIDOLS planning estimate).
• A tool switch lands firm-wide on one date, so everyone needs the basics on day one.
• Unsanctioned use is already common: data rules cannot wait for a cascade. Rolling out Microsoft 365 Copilot? Fix SharePoint oversharing first.
• Nobody owns data questions yet. Fix that first: no champion should decide what may be pasted.

Common failure modes

• The volunteer badge: enthusiasts self-select instead of the people colleagues copy.
• The unpaid helpdesk: no scope, no hours, no refusal list.
• The champion without authority: a good demo and no change.
• Demo-only sessions: people leave with notes about somebody else's work.
• The vanity dashboard: seats and sign-ins reported as adoption.
• The charter with no end date: stepping down later costs goodwill.

They sit alongside the project-level causes in why most AI projects fail.

Champions can start with the free AI Fluency course: 25 written lessons across 7 units and 3 role tracks, no sign-up, and 228 minutes for one full path. Its L&D and People track holds the session design, champion cascade and measurement lessons this guide draws on, and it is vendor-neutral by design.

Paid enrollment adds guided builds, starter templates, instructor video, cohort office hours, a graded capstone, a certificate issued for a working build that was submitted and marked (never for time spent or lessons read), and roster and completion reporting for organizations. Individual enrollment is not open yet, and the course page has a waitlist; organizations can ask about a cohort through AI Fluency training for teams.

Unsure which workflow your champions should start on? Begin with the free workflow map, or book a 15-minute call and bring one workflow that costs your team time. If you need a vendor's product certification or a course that claims to make you compliant, AIDOLS is the wrong fit: our certificate attests to a working build, nothing more.