Fourteen months after the kickoff, a transformation director opens the channel called #ai-champions to check something, and finds the last message was posted eleven weeks ago by her own program manager. It says "Happy Friday, everyone! Any wins to share?" and it has one reaction, a thumbs-up, from a person who left the company in March. The channel has 60 members. The slide that created them said "AI Champion Network: 60 volunteers across 9 functions" and it drew applause in the steering committee, because a number that big looks like momentum. Nobody asked the only question that mattered: what does a champion do on a Tuesday, and who released the hours for them to do it? The answer was nothing and nobody. The network was named, not built. And when the biggest deployment wave hit and the floor needed someone standing on it, there was one overwhelmed project manager, a rising queue of small unanswered questions, and by week three a set of workarounds that would take two years to unpick.

The List That Decayed, and Why It Was Always Going To

The champion network is the highest-leverage structure in enterprise AI enablement and the most reliably mismanaged. Both halves of that sentence deserve slow attention, because the gap between them is where most change budgets quietly die.

Start with the mismanagement, since it is nearly universal. The standard pattern is a volunteer enthusiasm list. Someone circulates a sign-up form, the people who like new tools raise their hands, their names go on a slide, and they are thanked warmly in the next town hall. Then they are given nothing: no protected hours, no authority to decide anything, no curriculum, no channel that reaches a decision, and no path that leads anywhere in their own career. What happens next is not a mystery, it is arithmetic. A champion has a full-time job with a full-time backlog, the champion work is unfunded, and unfunded work loses to funded work by about week three, every single time. No personality type is immune. The list decays within two quarters, the channel goes quiet, and the program writes the postmortem line that appears in a thousand decks: "we could not sustain champions." The truthful version is that nothing was ever built to sustain.

Now the leverage half, which is the reason this failure is expensive rather than merely embarrassing. A champion is the only enablement resource that is physically present at the moment of friction. Not in a scheduled session, not in a ticket queue with a four-hour response target, but at 10:40 on a Tuesday when a colleague two desks away stares at an output that looks wrong and has ninety seconds to decide whether to investigate it or quietly go back to the old spreadsheet. Everything your program spent money on (platform, redesigned workflow, verification gate, training curriculum) is decided in that ninety seconds, repeatedly, across hundreds of desks. Central enablement cannot be there. A champion already is.

Champions also carry two things no central program can manufacture. The first is dialect: a champion says "you know how we always have to re-key the vendor reference before it will match?" and is understood instantly, because they do the same job. The second is credibility, which is not purchasable, not transferable, and not delegable. When a respected colleague says "I use this and it is better," a room updates in a way it never does when a program manager says the identical sentence with better slides.

Here is the finding that reframes the task. MIT's 2025 GenAI Divide research found that while 95 percent of enterprise generative AI pilots delivered no measurable profit-and-loss return, a shadow AI economy thrived underneath: employees using personal AI tools for real work, learning by themselves, producing value that never reached a program dashboard. If your own shadow-AI census returned a list of self-selected, self-taught practitioners, the champions already exist. You are not creating a capability from nothing; you are performing four verbs on people already there: find, equip, protect, retain. Every failure in this lesson is a failure of one of those four, usually protect.

A champion list is enthusiasm. A champion network is an institution. The difference is hours, curriculum, authority, and a path, and only the first one costs money.

The artifact you build in this lesson is the Champion Operating Model: one page that states the four roles the network performs, the selection criteria, the protected-time contract, the equipping curriculum, and the retention mechanics. It is short by design. Its power is that every line commits somebody to something, which is precisely what the sign-up form never did.

The Four Roles a Champion Network Performs

Most programs treat "champion" as a single undifferentiated thing: a helpful person who likes AI. That vagueness causes nearly everything that goes wrong afterwards, because you cannot select for a vague role, train for it, fund it, or measure it. Naming the roles turns a list into an institution. There are four, and each implies different selection, equipping, and measurement.

Role one: floor support

This is the week-one-to-four presence funded in your change plan: someone visibly available on the floor (physical or virtual) during each wave's launch window, whose whole job in those hours is answering the small question before it becomes a workaround. The small question is the point. "It gave me two dates and I do not know which to use." "Do I still fill in the old form?" "Is it supposed to look like this?" None is worth a support ticket, none will surface in a survey. Left unanswered for three days, each becomes a private rule, and private rules aggregate into the shadow process your redesigned workflow was supposed to replace.

Floor support is the highest-volume role and the one that most obviously requires protected hours, because it demands availability at unpredictable moments. It also generates the raw material for role three, so treat every floor-support hour as producing two outputs: an unblocked colleague and a logged observation.

Role two: coach

This is the practice supervision from the training blueprint, where the ratio of one coach to six learners is not a preference but a planning constraint that sets your calendar. Coaching is where behaviour actually changes, because watching someone do the work is the only way to discover what they actually do, as opposed to what they said in a survey they do.

Coaching requires two things. The first is verifier-tier skill: a coach must be able to judge whether an output is acceptable against the gate standard, which is why coaches take the full 14-hour verifier curriculum rather than the shorter general track. The second is harder and routinely underestimated: the ability to watch someone work without taking the keyboard. Most technically strong champions are initially bad at this. They see the learner floundering, feel the discomfort, reach over, fix it in nine seconds, and feel helpful. The learner has watched a demonstration, the least effective form of instruction available, and has mainly learned that they are slow. Restraint under that discomfort is trainable and must be trained explicitly, which is why "great at the tool" is not a sufficient qualification.

Role three: signal carrier

This is the role that alone justifies the network's entire cost, and it is the one most often left unstructured. Your champion network is the program's sensory organ. Champions see, weeks before any dashboard, the things that decide whether the deployment survives: the workaround forming in a corner of the team, the screen taking four clicks where the redesign assumed one, the confidence-band mismatch where people trust outputs they should question and question outputs they should trust, and the quietly decisive one, the manager who has started allowing the old path "just while we are busy."

Unstructured, this is anecdote: a good story told at a coffee machine to someone who cannot act on it. Structured, it becomes a friction stream feeding your improvement loop. The structure is unglamorous: a standing monthly signal call, a simple intake with three fields (what you saw, where, how often), and a disposition process answering every signal with one of four outcomes: changed, scheduled, declined with a reason, or escalated. The disposition is the part that matters. A champion whose signal produced a visible change becomes a permanent asset; a champion whose signals disappear into a form becomes a former champion.

Role four: local credibility

The fourth role is the one you cannot buy, cannot schedule, and can destroy in a single meeting. It is the peer who says "I use this and it is better" in a room where the central program has no standing. That sentence is worth more than an all-hands, a redesigned intranet page, and a budget line, combined, because it comes with the speaker's own reputation attached.

Which produces the sharpest rule in the operating model: never ask champions to sell. The moment a champion defends something they do not believe (a rollout that is not ready, a claim about job security nobody has verified, a tool that is genuinely worse for their team's edge cases) they spend the credibility that made them valuable, and they do not get it back. They also, reasonably, stop returning your messages. Ask them to tell the truth, then take on the real work: designing the program so the truth is favourable. If your champions cannot honestly say the new way is better, you do not have a champion problem. You have a design problem wearing a champion problem's clothes.

The Artifact: The Champion Operating Model

One page, five parts. Write it before you recruit, because every part changes who you would recruit.

Part one: selection, or influence over enthusiasm

The criteria that mislead are seductive because they are easy to observe. Technical enthusiasm is visible: this person tries every new tool, posts about them, volunteers first. Tool proficiency is measurable. Job title is on the org chart. None of these predicts whether a champion will change anyone's behaviour.

The criterion that matters is peer influence, and it has a specific diagnostic question. Do not ask a team "who is your AI expert?" Ask: who do you check with before you change how you do something? The names that come back are the informal spine of the organization, and they are frequently not the names on the volunteer list. The enthusiast with no followers is a poor champion and often a poor coach, because the same traits that produce early adoption (comfort with ambiguity, tolerance for broken things, delight in figuring it out alone) produce impatience with a colleague who needs the third explanation.

The best champion in the building is a different creature entirely: the respected sceptic who has become convinced. Their conversion is evidence, because their peers know they were not easy to convince. When that person says the verification gate is worth the four minutes it costs, the room believes it in a way it never believed the program. So recruit converts deliberately: find the credible sceptics, give them early hands-on access, let them try to break it, and take their objections seriously enough to fix some. A sceptic whose objection produced a change is the most persuasive advocate you will have, and you cannot fake the process that produces one.

The second principle is coverage over count. Forty-one champions concentrated in two functions is a broadcast tower with no receivers: an impressive slide number and zero presence in the functions that have none. Map the list against the affected teams in your wave plan before celebrating the total. Every function with a wave slot and no champion gets targeted recruitment before the wave, not during it, because a champion recruited in week one of a launch is untrained during exactly the weeks that decide everything.

The third principle is a ratio derived from arithmetic rather than volunteer supply. Illustratively, one champion per 12 to 20 people in affected teams during an active wave, tapering after. Do not take that band as a standard; derive your own from two demands you already know. Floor support: how many hours of visible presence does a launch week need, against the hours per champion you funded? Coaching: how many learners need practice supervision at one coach to six? The larger number sets your requirement. If it exceeds the volunteers available, that is a finding for your sponsor, not a reason to stretch six people across nine teams and call it a network.

Part two: the protected-time contract

This is the make-or-break mechanic, and it is the reason most champion networks fail while their design documents look fine. Protected time is not an intention, an encouragement, or a bullet in a kickoff deck. It is a three-way written agreement between the champion, the champion's manager, and the program, and it contains four things:

• The hours and the window. N hours per week for M weeks, tied to a specific wave, with an end date. Open-ended commitments are refused by good managers and ignored by bad ones. A bounded ask is signable.
• What the hours are for. Named against the four roles: floor-support presence, coaching sessions, the monthly signal call, preparation. Unnamed hours evaporate into whatever is loudest.
• What the team gets in exchange. This is the clause programs skip and managers care about most. If a team lead releases six hours a week, something must give: a target adjusted, a queue reprioritized, a reporting task paused, cover arranged. Write the specific relief down. "We will absorb it" is how a champion ends up doing their day job plus champion duties and resenting both.
• The manager's signature. The critical one. The champion's willingness was never the constraint; their manager's Thursday afternoon capacity crunch is. A signature converts an ambiguous favour into an agreed allocation that both parties can point to under pressure.

Track refusals rather than burying them. A manager who declines to sign is telling you something true: either the team is genuinely at capacity (a fact your wave plan needs) or the manager does not believe in the deployment (a fact your change plan needs more urgently). Both belong on the risk register while there is still time to act.

Part three: equipping, a real curriculum

Champions are not equipped by a welcome kit. They need a curriculum, and it has four components, delivered before their wave rather than during it.

The artifacts component is the one programs forget, and its absence is visible from the floor within days. A champion who cannot explain why the verification gate exists will be asked to justify it by an irritated colleague on day two, and will either invent a justification (creating a rumour you will spend a month correcting) or shrug (teaching the room that the gate is arbitrary and therefore optional). Walk them through the real thing: the gate specification and its entry criteria, the override log and what happens to an entry after it is written, and the threshold that triggers escalation along with the named person it goes to.

Honest talking points means exactly that, including the current answer to the job question. If the answer is "no reduction in force is planned in this wave, and here is the redeployment commitment in writing," champions get that sentence and its date. If it is genuinely uncertain, they get the honest uncertainty and the date it will be resolved. What they must never be given is a reassuring line that turns out to be false, because the person who repeated it in good faith to their own team is the person you have just destroyed.

One cheap and disproportionately powerful mechanic: champions must never be the last to know a program fact. Every announcement, change of plan, or awkward decision reaches the champion network 24 hours before it reaches everyone else, with a short plain explanation. It costs one extra email and it buys two things: champions who look informed in front of their peers (which is most of what status means at work) and an early warning system, because their reaction to the message tells you exactly how the message will land.

Part four: retention and path

Networks die for one reason: after the novelty, there is nothing in it for the champion. The extra work is real, the recognition is a monthly newsletter, and the day job never got smaller. Five mechanics, in rough order of power:

1. Recognition tied to verified improvement. Not "thanks for your enthusiasm" but "this team's rework rate fell from 12 percent to 5 percent and this person coached the change." Recognition that names a number is credible and useful to the champion's own manager at review time.
2. Skills written into role descriptions. Verification, coaching, and process improvement are promotable capabilities. Get them into the job architecture so being a champion visibly builds a career rather than borrowing from it. This is the mechanic that turns a favour into an investment.
3. First access. Champions get new capability first, always. It costs nothing, it is genuinely useful for testing, and it converts the role into a privilege.
4. A visible community with real influence. A monthly call where champions see the disposition of their own signals: what changed, what is scheduled, what was declined and why. Champions who see their signals change the product stay. Champions who feel surveyed leave, and they leave quietly, by simply becoming busy.
5. Honest rotation. Define a term (two waves, or six months) with a named alumni role afterwards. Counterintuitive as a retention tool, and it works: a bounded ask is easier to accept and to renew, exit becomes graduation rather than burnout, and a champion who left honourably still says good things in rooms you are not in. Open-ended commitments end in silent attrition, costing you both the champion and the story they tell.

Part five: measuring the network itself

You measure the deployment. Almost nobody measures the network, which is how a program discovers in month six that it has been reporting a number that stopped being true in month two. Four measures, none expensive:

• Active versus nominal champions. Nominal is the list. Active means they did something observable in the last 30 days: a coaching session, floor hours, a signal raised. Report both numbers, always together. The gap between them is your program's honesty indicator.
• Signal volume and disposition rate. How many signals came in, and how many produced a change. The disposition rate is the number that proves the network is heard rather than tolerated, and publishing it back to the network is itself the retention mechanic.
• Coverage against the map. Champions per affected team per wave, with the gaps named. A single red cell here is worth more attention than the headline total.
• Champion-adjacent adoption. Adoption and quality metrics for teams with a champion versus teams without one. Waves rarely have perfect coverage, which hands you a natural experiment for free. Measure it, report it with its caveats (teams differ in more ways than one), and you hold the most persuasive artifact available when someone proposes cutting protected hours to save budget.

Worked Example: From 41 Opt-Ins to 18 Active Champions

All numbers here are illustrative, built to show the arithmetic rather than to be copied. The setting is the programme running through this chapter: a shadow-AI census that produced 41 opt-in volunteers, a wave plan, and a change plan funding floor support at four champions per wave at six protected hours a week.

Triage. The strategist does not accept 41. She runs the influence question through 14 team leads ("who does your team check with before changing how they work?") and cross-references the answers with the opt-in list. The overlap is partial and instructive: of the 41 volunteers, 22 are named by nobody, and nine names come back that never volunteered at all. The triage lands on 18 active champions for waves 0 to 2: six high-influence converts, nine respected practitioners, three approved candidates whose managers declined the time. The six converts include an accounts-payable team lead who spent the discovery workshops arguing that the invoice-matching redesign would not survive contact with real vendors, was invited to try to break it, found two genuine failure cases that were fixed, and is now the most quoted person in the finance function on the subject. That arc, sceptic to tester to advocate, is the recruitment pattern to industrialize.

Manager refusals. The three declines are escalated to the wave sponsor with the specific ask (six hours a week for four weeks, with named relief). Two are resolved: one by moving a monthly reporting task off the champion for the wave duration, one by shifting the champion into the next wave, after the team's peak season. The third refusal stands, and the honest reading is genuine capacity, which goes onto the wave-2 risk register as a coverage gap rather than being smoothed over.

Coverage and targeted recruitment. The map shows the real problem the total was hiding. Two functions have zero opt-ins: sales, which the readiness audit scored at 11 percent capability and is effectively greenfield, and HR. No volunteers means no list to triage, so the strategist uses the route that works where volunteers do not exist. She asks those functions' managers the influence question directly, gets five names, and approaches each one personally with a specific offer: this is the role, these are the hours, this is what your team gets in exchange, this is the training, this is the term. Three accept, and they are equipped before the wave-3 slot rather than during it. In low-capability functions nobody volunteers, because volunteering requires already feeling competent. Those functions need to be asked, individually, by name.

Protected time. Fifteen managers sign the three-way agreement: six hours a week for four weeks per wave, named against floor support, coaching, and the monthly signal call, with the exchange clause completed in every case. Three agreements specify a paused reporting task, four specify a reduced target, the rest specify cover arrangements. The total commitment is roughly 360 champion hours across the first two waves, which is a real number the program can defend and a real number the finance partner can see.

Equipping. Every champion completes the 14-hour verifier tier, the 4-hour coaching module, and the artifact walkthrough covering the gate specification, the override log, and the escalation path. The feedback is consistent: the observation exercise (watch a colleague work for ten minutes without speaking) is the most uncomfortable and most valuable hour of the curriculum.

Results at week 12. Sixteen of 18 champions still active, with two rotated out honestly at the end of their term and moved to the alumni role rather than drifting into silence. Forty-seven signals raised, of which 19 produced a change: seven interface or process fixes, five clarifications to the gate specification where the standard turned out to be ambiguous, four training-content revisions, three escalations about manager behaviour. The disposition report is published back to the network in full, including the 28 signals that did not produce a change and the reasons, and the signal volume in the following month goes up rather than down, which is exactly the behaviour a heard network shows.

The comparison nobody planned for is the most useful output. Championed teams reach 84 percent adoption of the redesigned workflow by week 12. The one wave-1 team with no champion available (the manager refusal) reaches 61 percent, with a higher override rate and two documented workarounds. Reported honestly, with the caveats that this team also had the tightest seasonal workload and that one team is not a controlled trial, the 23-point gap is still the most concrete evidence the program has about what protected champion hours buy. When the next budget cycle proposes trimming those hours, that gap is the answer, and it exists only because someone thought to measure it.

The Failure Story: Sixty Ambassadors and a Newsletter

A retail group with 4,000 employees launches an AI assistant across merchandising, store operations, and customer service. The change lead, who has read the right things about peer influence, proposes a champion network. Leadership approves enthusiastically, because it is the cheapest item in the plan: a sign-up form goes out, 60 people volunteer, and they are christened AI Ambassadors. A welcome kit arrives with a branded water bottle and a "top ten prompts" card, and a monthly newsletter begins. The slide says 60 ambassadors across 9 functions and is, briefly, the most popular slide in the deck.

Nothing else is provided. No hours are released, so ambassador work happens in the gaps, which in a retail operation means it happens never. No curriculum, so ambassadors know roughly what a colleague who watched the same 45-minute webinar knows and cannot answer the second question about verification. No channel reaching a decision: the newsletter goes out, nothing comes back. No term, no path, no recognition beyond a name in a list.

The decay is textbook and forecastable on day one. Month two: eleven ambassadors still posting. Month four: the channel is quiet apart from program announcements. Month six: a pulse survey finds most ambassadors have not used the tool in six weeks, and three do not remember volunteering. Meanwhile the largest wave, 600 store colleagues, launches with floor support consisting of one project manager covering 40 locations by video call. Week-two friction goes unresolved: two questions about which system remains the record of truth, one screen requiring re-entry of data the tool already has. Within a month those are established local practice, and by the time anyone measures it, the redesigned workflow exists on paper while the floor runs a hybrid of old and new that is slower than either.

The postmortem blames change fatigue, the corporate synonym for "we did not fund the change." The honest finding is narrower: the network was never built, only named. Sixty people said yes to a question that was never followed by an offer. Enthusiasm is a real input and it is not a plan; an institution is what converts enthusiasm into capacity, and an institution is made of hours, curriculum, authority, and a path. BCG's 10-20-70 rule (10 percent of AI effort in algorithms, 20 percent in technology and data, 70 percent in people and process) is not an inspirational statistic, it is a budget instruction, and the champion network is one of the places the 70 percent physically goes. McKinsey's finding that 88 percent of organizations use AI regularly while only about 39 percent report any earnings impact describes, at scale, the difference between naming and building.

What to Do Monday Morning

1. Triage your opt-in list by influence, not enthusiasm. Take the volunteer list you have and ask eight to fifteen team leads the influence question: who does your team check with before changing how they work? Cross-reference, expect partial overlap, and treat both surprises (volunteers nobody named, names that never volunteered) as findings.
2. Map coverage against your wave plan and name the empty functions. Put champions per affected team in a grid. Where a function has a wave slot and no champion, especially a low-capability function where nobody volunteers, go and ask specific people directly, with a specific offer, before that wave.
3. Get the three-way protected-time agreement signed. One page: hours, window, what the hours are for, what relief the team gets in exchange, and the manager's signature. Escalate every refusal, and log the ones that stand as coverage risks rather than letting them disappear.
4. Build the equipping curriculum from your own artifacts. Verifier tier plus a coaching module plus a walkthrough of your actual gate specification, override log, and escalation path, plus honest talking points including the current, dated answer to the job question.
5. Start the signal loop and publish the first disposition report. Book the monthly call, create a three-field intake, and within 30 days publish back to the network what came in, what changed, what is scheduled, and what was declined and why. That single report does more for retention than a year of newsletters.
6. Set the term and the alumni role now. Decide the length before you recruit, so the ask is bounded, the renewal is a choice, and the exit is a graduation.

Key Takeaways

• Treat champions as found rather than created: MIT's shadow AI economy finding means self-selected, self-taught practitioners already exist in your organization, and your four verbs are find, equip, protect, and retain.
• Name the four roles the network performs (floor support, coach, signal carrier, local credibility), because a single vague "champion" role cannot be selected for, trained for, funded, or measured.
• Structure the signal-carrier role deliberately: a monthly signal call, a simple intake, and a published disposition of every signal turns anecdote into the friction stream that feeds your improvement loop, and this role alone justifies the network's cost.
• Never ask a champion to sell something they do not believe; ask them to tell the truth and do the harder work of designing a program where the truth is favourable, because peer credibility is unpurchasable and non-renewable.
• Select on peer influence over technical enthusiasm by asking who people check with before changing how they work, and recruit converted sceptics deliberately, since their conversion is the evidence their peers actually weigh.
• Prioritize coverage over count and size the network from arithmetic (floor-support hours and the one-to-six coaching ratio), illustratively one champion per 12 to 20 affected people during a wave, rather than from whoever volunteered.
• Sign the three-way protected-time contract with the champion, the manager, and the program, specifying hours, purpose, and the relief the team receives, because an unfunded champion loses to their day job by about week three, every time.
• Retain through recognition tied to verified improvement, promotable skills in role descriptions, first access, real influence over decisions, and an honest term with an alumni role, and measure the network itself: active versus nominal, signal disposition rate, coverage, and championed versus unchampioned adoption.