AI champion

What is an AI champion?

An AI champion is somebody inside a team who has a formal part of their week set aside to help colleagues use AI well. They answer the small questions, they show what worked on the company's own work, and they carry the awkward cases back to whoever owns the AI usage policy. It is a named role rather than a job title: a slice of an existing job, written down. The person is deliberately not in IT, because what makes the role work is a seat among the people doing the work.

Microsoft's adoption guidance for Power Platform describes champions as the people their peers already go to, who build and share knowledge even when it is not part of their official job role. That last part is where most programmes fall over: if helping colleagues is not in the job, it is the first thing that goes in a busy week. The role is how a small company buys a piece of the seventy percent in the 10-20-70 rule, and in most SME AI plans it is the only budget line that goes to people instead of to software.

The questions people actually ask

People ask a colleague things they would never raise on a training day. The questions are not "what is a large language model". They are "can I paste this customer's quotation in here" and "it wrote a better version of my mail than I did, am I allowed to send that".

They arrive at the moment of the work, on a Tuesday in March, not on the day the training was booked, and several of them cost something to ask out loud, because asking means admitting you did not understand the tool. People will ask the person at the next desk what they will not put their hand up for in a room of twenty, which is why a champion sees your shadow AI before anyone else does.

A champion network versus a training session

The dimension that separates the two is when the help arrives relative to the question.

• A training session arrives before the question. The date comes out of a calendar, so the examples are generic, and asking costs something because you ask in front of colleagues. It does one thing well: twelve people get the same vocabulary in one morning.

• A champion arrives after the question, usually the same day. The example is the document the person is holding, and asking costs a walk to the next desk. What you get is the case nobody anticipated.

Neither replaces the other, and Microsoft's adoption material makes the same point from the other side: learning from co-workers is among the most effective ways people learn at work.

The four things that decide whether the role works

1. Time that is allocated and defended. Two hours a week, in the calendar, and a team leader who protects them in the week the orders pile up. Microsoft's guidance tells organisations to decide the level of commitment and time investment up front, communicate it, and get manager approval where needed. If you cannot say how many hours, you have found a volunteer, not created a role, and a programme that names twelve of them and gives none an hour is a distribution list within six months.

2. One named person to escalate to. Not a mailbox and not "IT". Somebody who owns the AI usage policy and can decide. "A customer has told us not to run their files through AI" is not a question a colleague should answer alone on a Friday afternoon.

3. Permission to say a use is a bad idea. A champion who is only allowed to say yes is doing marketing. The role earns its keep the first time somebody says "do not put the tender in there" and management backs them where everyone can see it.

4. Something to show that is not attendance. Three prompts other people now reuse, one rewritten work instruction, and a page with the five most frequent questions and their answers. That is the output, and it is the part that survives when the champion changes job.

Who to pick

Select for curiosity and for credibility with peers, not for technical skill. The most technical person on the team is often the wrong choice, for a mechanical reason rather than a social one: their questions are not their colleagues' questions. They passed the confusing stage months ago, so they answer about how the model works when the question was whether the customer's contract may go into it. Microsoft's guidance says most organisations spot champions rather than call for volunteers: watch who already answers questions in the team channel, and invite that person privately.

Two hours a week in a company of thirty

Here is where those two hours go across one month, for a champion in an operations team of eight. Eight hours in total.

• Three hours of questions at the desk. Roughly a dozen. Most take under ten minutes, two or three take half an hour because the answer is "show me the file".

• Two hours on one recurring job. This month it is the mail that goes out when a supplier delivery slips. The champion writes a prompt using three real past mails as examples, tests it against last week's delays, and puts the working version where the other seven can copy it.

• One hour of showing, split over two Mondays. Twenty minutes at the team meeting with the real mail and the real prompt on screen, not a product demo.

• One hour on two awkward cases. Somebody pasted a signed customer contract into a free chatbot. Somebody else wants to use AI to draft an assessment of a temporary worker. Both get written up in five lines and handed to the person who owns the policy.

• One hour writing it down. At the end of the month there are seven answers on one page, and four of the dozen questions never have to be asked again.

Two hours a week is around ninety working hours a year. At an internal loaded cost of 45 euro an hour that is a bit over 4,000 euro, and almost none of that lands on an invoice. For a company of thirty it is the cheapest thing in the AI budget that changes what somebody does on a Wednesday morning, and we would spend it before spending on more seats.

What the AI Act asks about AI literacy

Article 4 of the EU AI Act applies to providers and deployers of AI systems and has applied since 2 February 2025. A company using ChatGPT or Copilot for its own work is a deployer. The original wording required them to ensure, to their best extent, a sufficient level of AI literacy among their staff and anyone operating AI systems on their behalf.

That wording has changed. The Digital Omnibus on AI, Regulation (EU) 2026/1744, was published in the Official Journal on 24 July 2026 and entered into force on 27 July 2026. It rewrote Article 4 into a duty to take measures that support the development of AI literacy, and added that this does not require a provider or deployer to guarantee any specific level of AI literacy for any individual. The duty did not disappear. It became an obligation of effort rather than of result.

There is no certificate, no exam and no mandatory course. The European Commission's questions and answers on AI literacy say an internal record of trainings and other guiding initiatives is enough, and that copying a practice out of its public repository does not by itself grant a presumption of compliance. A champion network is one of the measures you can put in that record, next to a written AI usage policy. No article names the role, so do not create it in order to satisfy Article 4. Create it because the questions are being asked anyway. Article 26 is a separate matter: a deployer of a high-risk system has to assign human oversight to people with the necessary competence, training, authority and support, and a part-time peer is not that.

What to watch out for with an AI champion

The unpaid helpdesk. The questions do not stop when the two hours do. Six months in, one person is fielding everything and doing the rest of their job after five. Microsoft's guidance puts it plainly: champions are not a support team, they are business representatives. Write the repeated answers down and send tool problems to whoever owns the tool, because quietly adding hours only moves the ceiling. This is where AI fatigue starts, in the person you can least afford to lose.

Enthusiasm ahead of the policy. The champion who spent Sunday connecting a model to the shared drive has created the citizen developer problem with a shorter fuse. Fix that with the escalation name and the written policy, not by damping the enthusiasm that made the role work.

Dependency on the person. A team at the University of Melbourne published a study in July 2026 of three Australian organisations that had kept the same early-childhood education programme running for ten years or more. The sector is not ours, but the mechanism is about where knowledge ends up when one person is very good at holding it, which travels. In all three, expertise collected in the individual instead of in the organisation, colleagues leaned on that person for quality checks instead of on any system, and the champion being effective hid that nothing had been written down. The guard is the written output from the fourth point above, plus a second person doing the role beside them.