AI champions: how to build the internal network that makes training stick

What we mean by an AI champion

This post is about champions inside a company. The UK government's AI Champions scheme is unrelated. Here, an AI champion is a respected colleague who has already made AI part of how they work. They help the people around them do the same. Their job is translation: turning generic capability into "here's how this helps you draft that report, triage those tickets or prep that deck." It isn't a job title, and it isn't another name for the resident engineer.

The common mistake is to appoint the most technical person on the team. They're often the worst choice, because their habits and instincts don't match how everyone else works. The champions who move a team are usually mid-level people with real, unglamorous use cases. They have enough credibility that colleagues will admit to them what they don't understand. We make the same case for staffing an AI centre of excellence with domain experts rather than engineers.

Why champions matter: the tools are rarely the problem

Most organisations already have the tools. McKinsey's State of AI shows adoption is near-universal while measurable impact stays rare. The bottleneck has moved from access to use, and buying more licences won't clear it.

MIT's GenAI Divide report found that roughly 95% of enterprise generative-AI pilots produced no measurable P&L impact. It put most of the failures down to a learning gap rather than model quality. Generic tools were dropped into specific workflows, and neither side adapted. Training closes part of that gap, and champions stop it reopening after the course ends. We unpack the failure pattern in why 95% of AI pilots fail.

Without reinforcement, the skills from a one-off course fade by Friday. A champion is that reinforcement: the person two desks away you can ask when you're stuck. They'll walk you through it on your own tasks rather than on a slide.

Who makes a good champion

Recruit for trust and judgement over deep tooling knowledge. Look for people who are:

• Credible with their peers. Colleagues already go to them for help and take their word seriously.
• Already using AI on real work. Curiosity is a start, but you want a track record, however small.
• Generous with their time and patient with beginners. A champion who makes people feel stupid sets adoption back.
• Close to the day-to-day work. They should feel the same friction their team feels.
• Willing to say when AI is the wrong tool. Champions who oversell lose trust fast, and trust is all a champion has.

Notice that none of these says "most technical". A champions network runs on trust first and skills second.

How to set the programme up

Appoint a champion, leave them alone, and they'll quietly disappear. The programme is the support you build around them. It should also sit inside a wider adoption plan, like the one we set out in how to roll out AI across your team. In practice, that means:

• Pick a sensible ratio. One champion for a 200-person org is a mascot. A workable starting point is one champion for every dozen or two people, weighted towards the teams doing the most AI-exposed work.
• Train the champions first, and do it well. They can't lead what they haven't practised, and this is where expert-led training earns its place. We weigh up the options in in-house vs outsourced AI training.
• Protect their time. Bolt "champion" onto a full workload with no allowance and it evaporates within a month. Ring-fence a few hours a week, and say so out loud.
• Anchor it to the team's own work. Champions should run clinics and office hours on live tasks rather than demo files.
• Connect them to each other. Give them a shared channel and a regular session to pool prompts and warn each other about what went wrong. Isolated champions burn out.
• Give them air cover. That means a visible mandate from leadership and a clear line on what's approved and what isn't. Otherwise champions end up inventing governance on the fly.

Where external training fits

You don't have to choose between champions and outside training. They do different jobs at different stages. Hands-on training comes first, to set the standard and get a first cohort fluent. Champions then take over the day-to-day habit and keep it going between sessions. We scope our AI training to fit that, whether it's one team or a full rollout.

The programme we delivered with a training partner for a Fortune 500 tech company followed that shape. It ran as six weeks of live online sessions with about twenty senior product managers and UX designers. Each week was built on the cohort's own work rather than canned exercises, so habits got tested on real tasks before the next session. A cohort trained that way is where your first champions come from, because they've already done the reps.

How to tell if it's working

Judge the network by what it's still doing months after the launch. Attendance and a satisfaction score prove a day happened, and not much else. For the bigger question of whether the training itself worked, see how L&D leaders measure what stuck.

Better signals: are colleagues bringing real tasks to champions without being asked? Is the number of people using AI on core work rising month on month? Are champions spotting workflows worth standardising across the team? Take a baseline before you start, because you can't show movement from a number you never wrote down. Our guide to measuring AI ROI covers how to do that without fooling yourself.

When the support is in place, a champions network is the cheapest and most durable way to keep adoption going. If yours goes quiet after a month, look at that support before you blame the people. It's almost always one of three gaps: no protected time, no mandate worth the name or no community to lean on.