Most champions networks are announced with a list of names and no job description. Three months later the names are still there and nothing is happening.
The idea behind a champions network is sound and the research supports it, modestly. Peer influence moves behaviour by roughly ten percentage points, which is real but is not transformation. What kills most networks is that nobody writes down what a champion actually does, so the role defaults to unpaid technical support and the person burns out somewhere around week eight. This piece covers the six-field role definition, the selection questions that predict whether someone will still be doing this in month six, how to protect the time in a way that survives a busy quarter, and a measurement approach that looks at whether work changed rather than whether the network held its meetings.
Say an operations coordinator puts her hand up in March. The company is rolling out Copilot, someone asks for volunteers to be AI champions, and she’s the person on her team who’d already been using it for meeting notes. Her name goes on a slide. There’s a kickoff call with fourteen other champions and a shared Teams channel.
By June she’s getting something like thirty or forty messages a week. How do I get it to read this spreadsheet. Why did it make that up. Can you look at this before I send it. She answers them, because she’s helpful and because nobody else is going to. Most of it happens after five o’clock, because that’s when she has time. Her own manager has never mentioned the champion thing at all, and her objectives for the year don’t reference it.
Around September she stops replying quickly, and then she stops replying at all. The channel goes quiet. Nobody announces that the champions network has ended, because nobody ever announced what it was supposed to do in the first place.
That shape is common enough that it’s worth being precise about what went wrong, because the underlying instinct was right. Training builds capability: someone can now do the thing, in principle. Workflow design gives that capability a real recurring task to land in. Reinforcement is what turns it into behaviour that survives the second week. A centralised training programme can do the first one well and struggles with the other two, because it isn’t in the room on a Tuesday when someone is stuck on their own actual spreadsheet. A champion is.
The evidence for peer influence is real and it’s smaller than most people assume. A Cochrane review pooled 24 randomised trials of using respected local peers to spread new practice, and found a median absolute improvement of 10.8% in colleagues’ compliance. [1] The interquartile range ran from 3.5% to 14.6%, and for some comparisons it included the possibility of a small negative effect.
So roughly ten points. Worth doing, not a transformation, and highly variable depending on how it’s run. Which raises the obvious question of what separates the trials at 14% from the ones at 3%, and here the review says something uncomfortable that maps almost exactly onto corporate champions programmes: “In most studies, the role and actions of the OL were not clearly described, and we cannot, therefore, comment on strategies to enhance their effectiveness.”
After 24 randomised trials, nobody had written down what the champions actually did, which is roughly the same problem most corporate programmes have. The difference between a network that shifts behaviour and one that fills a channel is almost entirely in how specifically the job was described.
A champion without a defined scope becomes a help desk.
Everything below is an attempt to write the role down properly, because that’s the variable nobody controls and it’s the one that decides whether you get the 14% or the 3%.
Here’s the test worth applying to your own programme. If you asked three of your champions to describe their role in a sentence, would you get three answers that resemble each other? In most companies you get “help people with AI,” which isn’t a role, it’s a direction of travel.
A champion role that survives has six fields filled in, and it fits on half a page. This is what one looks like when it’s done properly, for a marketing operations coordinator:
| Field | Marketing operations champion |
|---|---|
| The work this is about | Campaign reporting and the weekly performance summary. Not “AI in marketing” generally. |
| What the champion does | Runs a 20-minute session on the first Tuesday of the month showing one real task done with AI, start to finish, using the team’s own data. Adds working prompts to the shared library. |
| What the champion does not do | Troubleshoot licences, access or logins. Answer questions about data policy. Fix anything outside campaign reporting. All three go to a named person instead. |
| The monthly output | One documented before-and-after: the task, how long it took before, how long it takes now, what still needs a human, and whether they’d recommend it to the rest of the team. |
| Protected time | Three hours a week, agreed in writing with their line manager, and reflected in their objectives. |
| Who they report to | Head of marketing ops, who reads the monthly write-up and replies to it. |
A worked example of the six-field role definition described in this section. Copy the field names, replace the right-hand column.
The two rows people leave blank are what the champion does not do, and the protected time. Those are the two that do all the work. Leaving out the first is how you get the help desk. Leaving out the second is how you get someone doing it at seven in the evening until they don’t.
The monthly write-up matters more than it looks. It’s the thing that converts a champion from a person who is enthusiastic into a person who has produced something, and it’s the only artifact you’ll have when someone senior asks whether any of this is working. It should take them about forty minutes, and the drafting part of it is a reasonable thing to hand to AI, as long as the split is explicit:
| What AI does | What the champion still owns | How it gets checked |
|---|---|---|
| Turns their rough notes and timings into a structured before-and-after in the standard format | The judgement call on whether it genuinely saved time, and the honest note on what it still gets wrong | Head of marketing ops reads it within a week and replies. If nobody replies twice running, the programme is already over |
The explicit split for the one piece of the champion role where handing work to AI makes sense.
There’s some evidence that teams doing this kind of open sharing are a different animal. Microsoft’s 2026 Work Trend Index found 61% of its most advanced AI users say their teams share tips, new agents, learnings and mistakes with each other, against 36% of everyone else. [2] That’s a survey, so it’s an association rather than proof that the sharing caused the capability. It’s still the closest thing to a description of what a working network looks like from the inside: not a channel where people ask for help, a channel where people post what they tried.
The default selection method is to ask for volunteers, and the default result is that you get the people who were already excited. That’s not nothing, and it’s also how you end up with a network of enthusiasts who have no influence over anyone at all.
The failure looks like this. Someone speaks up in the launch meeting, clearly knows more about AI than anyone else in the room, and gets picked immediately. Six months later their team still hasn’t changed how it works, because the reason they knew more about AI than everyone else is that they’re the person who tries every new tool and nobody quite takes their recommendations seriously. Enthusiasm was visible. Peer trust wasn’t, so nobody checked for it.
These five questions are worth scoring before you confirm anyone, and they take about ten minutes per candidate:
| Question | 0 | 1 | 2 |
|---|---|---|---|
| When this person recommends a way of working, do colleagues adopt it? | Rarely | Sometimes | Usually, and you can name an example |
| Do they own a recurring process the team depends on? | No | Partly | Yes, weekly or monthly, and they run it |
| Has their manager agreed to the hours, in writing? | Not asked | Verbally | Yes, and it is in their objectives |
| Do people already ask them things informally? | No | Occasionally | Yes, this is already happening unpaid |
| Can they explain something without making the other person feel slow? | No | Sometimes | Yes, and colleagues say so |
A scoring aid, not measured research. 8 to 10: confirm. 5 to 7: confirm only if the manager signs off on the time. Below 5: they may be a great AI user and the wrong champion.
The second question does more work than the others. A champion who owns a recurring process has somewhere to put the new way of working immediately, which is the difference between capability and behaviour. A champion who is enthusiastic but owns nothing recurring has to persuade someone else to change, and that’s a much harder job than the one you thought you were giving them.
Pick for peer trust, not for volume of enthusiasm.
It’s also worth being specific about which roles tend to work, because “pick an enthusiastic person from each team” hides a real difference. A coordinator or specialist who runs a recurring process is usually the strongest choice: close to the work, trusted, and with something to change on Monday. A team lead is a good second, but their demonstrations land differently, because when a manager shows you a faster way of doing something it can read as an instruction rather than an offer. A director or department head is usually the wrong pick for the champion role and the right pick for the sponsor role, which is a separate job: they clear the time, they read the write-ups, and they don’t run the sessions.
Numbers, since it’s the question everyone asks. One champion per team that has its own recurring workflows is the useful unit, not a percentage of headcount. Six champions covering six genuinely different workflows will do more than twenty covering the same three.
Three selection mistakes come up often enough to name:
The drift into technical support happens for an understandable reason. A champion is the most visible AI-related person nearby, and people route every AI-shaped question to the nearest AI-shaped person. Licence problems, password resets, “is it allowed to see this document,” someone’s laptop. None of that is the role, and all of it arrives anyway.
The fix is to publish the boundary rather than hoping people infer it, and to make sure the things outside the boundary actually have somewhere to go. A boundary with no alternative destination is just a champion saying no.
| Question that arrives | Who it belongs to | Where it goes |
|---|---|---|
| “How would you get it to summarise this campaign report?” | Champion | Team channel, answered in the open so the answer is reusable |
| “It’s making things up in my draft.” | Champion | Team channel, and it becomes next month’s session if it happens twice |
| “I can’t log in / I don’t have a licence.” | IT service desk | Normal ticket, linked in the channel description |
| “Am I allowed to put client data in this?” | Whoever owns the AI usage policy | Named person, not the champion, and not a guess |
| “Can you just do this one for me?” | Nobody | Champion shows them once, in a session, and the answer goes in the library |
Publish this in the channel description at launch. The last row is the one that quietly eats a champion’s week.
The support the champion themselves needs is easy to forget, since they’re the one doing the supporting. BCG’s 2025 global survey of employees found 84% of people with access to a coach were regular AI users, against 70% of those without. [3] Worth holding loosely, because BCG fielded and analysed it themselves and published no field dates or independent research agency, and because regular users may simply be the sort of people who seek out coaches. But the direction is consistent with the rest of the picture, and champions are the group most likely to be given a role and no coaching at all.
Practically, that means the sponsor is not a figurehead. Half an hour a month with the champions as a group, where they compare what’s landing and what isn’t, is the whole intervention. Teams run that internally all the time and it works fine. Where an outside facilitator earns their fee is in the first two or three sessions, when nobody yet knows what a good demonstration looks like and the group is still deciding whether this is real. Our corporate workshops usually start there, and then hand the cadence back.
Protect the hours in writing, or don’t create the role.
Month two is where these things die, and it’s predictable enough to plan for. The launch energy has gone, the novelty questions have been answered, and the champion is now doing an ongoing job that was sold to them as a short burst of helping out.
Three things keep it alive, and none of them are motivational.
On the time itself, there’s a useful piece of context. Slack’s Workforce Index, fielded in August 2024 across 17,372 desk workers, found 61% had spent under five hours in total learning to use AI, and 30% had had no training at all, including no self-directed experimentation. [4] That’s two years old now and adoption has moved a long way since, so treat it as a snapshot of how the starting position looked rather than a current number. It still frames the ask correctly. Three protected hours a week for a champion is not a modest request inside an organisation where most people have had under five hours, total, ever. It’s an unusual investment in one person, and it should be argued for as one.
The cadence that seems to hold up:
Champion answers in the open channel. Nothing scheduled, nothing formal.
One 20-minute session on a real task, and the written before-and-after to their sponsor.
Half an hour with the other champions. What landed, what didn’t, what to stop doing.
Renew, rotate or close. A decision either way, made out loud.
A recommended cadence based on how these programmes tend to fail, not a measured finding.
The tempting measurement is attendance. How many champions, how many sessions, how many people in the channel. All easy to pull and all describing whether the programme happened rather than whether it did anything.
The lens worth using instead has three stages, and most reporting stops at the first.
| Stage | The question | Where the answer comes from |
|---|---|---|
| Use | Did anyone try the thing the champion demonstrated? | Ask at the next session. Show of hands. Takes ninety seconds |
| Persistence | Are they still doing it a month later, without being reminded? | The champion asks two people directly. Not a survey |
| Impact | Is the actual work better, faster or different, and would anyone notice if it stopped? | The monthly before-and-after write-ups, read together at six months |
The Future Factors measurement lens applied to a champions network. Usage numbers can be the first signal, but on their own they measure access rather than adoption.
The persistence question is the one that tells you something you didn’t already know. Almost anything gets tried once after a good demonstration. The interesting number is how much of it is still happening in week five, when the person is busy and the old way is still available and nobody is watching.
Measure whether the work changed, not whether the network met.
At six months, read the write-ups as a set. You’re looking for one thing: workflows that are genuinely different now and would have to be actively undone to go back. If you have four of those from six champions, the network is working and worth renewing. If you have twelve enthusiastic sessions and no workflow that changed, you’ve run an internal events programme, and it’s better to say that plainly than to keep it going out of politeness.
This sits inside a bigger picture, and it’s worth knowing where the boundaries are. What people need in the first fortnight after a tool arrives is a different problem, covered in the AI user enablement guide. The people whose agreement you need before any of this starts are a different problem again, and that’s the stakeholder conversation. A champions network is neither of those. It’s the layer that keeps working after the launch attention has moved on, which is precisely why it needs a structure that doesn’t depend on anyone staying excited.
If you’re starting this week, do one thing. Take the six-field role definition, fill it in for a single champion covering a single recurring workflow, and get their manager to agree the hours in writing before you announce anything. One properly defined champion will teach you more about whether this works in your organisation than fourteen names on a slide.
A group of employees, usually one per team, who are given a defined role in helping colleagues actually use AI in their day-to-day work, rather than a general encouragement to be helpful. The working version has three things a name-on-a-slide version doesn’t: a specific scope tied to a real recurring workflow, a stated list of what the champion does not handle, and protected hours agreed with their own line manager. It sits alongside training rather than replacing it. Training builds the capability; the champions network is where that capability meets a real Tuesday afternoon task and turns into a habit.
Think in workflows rather than headcount percentages. One champion per team that owns its own recurring processes is the useful unit, and six champions covering six genuinely different workflows will produce more change than twenty covering the same three. The number people usually land on is too high, because it’s calculated from company size rather than from how many distinct kinds of work there are. A 300-person company with four departments doing four fundamentally different jobs might need five or six. The failure mode of too many is that none of them have a clearly distinct patch, so the role blurs and the write-ups start repeating each other.
Pay is rarely the deciding factor and protected time almost always is. If someone is being asked for three hours a week on top of a full workload, a bonus doesn’t help them; it just makes the impossible ask better compensated. What works is time that has been agreed with their line manager in writing, with something specific coming off their plate, plus recognition that reaches the people who decide their next role: the monthly write-ups referenced in their performance review, and one presentation to the leadership group. If your organisation genuinely can’t free three hours a week for one person per team, that’s useful information about how much appetite there really is, and it’s better to find out before the launch email than in month three.
The role is announced but never defined, so it defaults to unpaid technical support. Every AI-shaped question in the building routes to the nearest AI-shaped person, including licences, logins and data-policy questions that were never the champion’s job. Because none of it was written down, the champion can’t decline any of it without seeming unhelpful, and the work happens in the evenings until it stops happening at all. The two fields that prevent this are the ones most often left blank: an explicit list of what the champion does not do, and a named alternative destination for each of those things. A boundary without a destination is just a champion saying no, which is worse than no boundary.
IT support fixes things that are broken. A champion changes how work gets done, which is a different job requiring different evidence of success. IT can tell you whether someone has a licence and whether the tool is running; they can’t tell you whether the weekly campaign report is genuinely better now. Keeping them separate protects both. The moment champions start handling access and permissions, the demonstration work stops, because troubleshooting is urgent and workflow change never is. That’s why the scope table matters more than it looks: it isn’t bureaucracy, it’s the mechanism that stops the urgent work from eating the important work.
The four figures in this article come from their original publishers rather than from coverage of them, all checked on 29 August 2026: the Cochrane review abstract, the Microsoft Work Trend Index 2026 report itself, BCG’s own AI at Work 2025 publication, and Slack’s Workforce Index PDF including its methodology statement. Two caveats are stated in the body and worth repeating here. The Cochrane evidence is from healthcare rather than workplace AI, so it establishes that peer influence works and roughly how much, not that it works identically in your company. The BCG figure is self-fielded and self-analysed with no published field dates or independent research agency, and it is correlational. Two widely circulated statistics were deliberately excluded: a claim that 69% of employees learn AI primarily from colleagues, and a claim that peer-led adoption achieves 2.1 times higher sustained usage. Both are exactly what this article wanted and both trace back to a single content-marketing site citing unpublished vendor program data with no sample size, field dates or method. The role definition, selection scoring, scope boundaries and cadence are Future Factors’ own, developed from running adoption programmes, and are described that way rather than presented as research findings.
Weekly tips. Real examples. Practical help for busy professionals.
We care about your data, check out our privacy policy.