The steering committee meeting is in month five when it happens. The pilot's sponsor is halfway through a slide about adoption metrics when the head of Treasury clears her throat and says, quietly, that her team has "serious unresolved concerns about the control implications," and that she has discussed them with the CFO, and that perhaps the pilot should pause until a proper review is complete. Around the table, three other heads nod, a little too readily, in a way that tells you this conversation happened before the meeting, several times, in rooms you were not in. The transformation lead who owns the pilot feels the floor tilt. Nothing on any dashboard predicted this. Usage was fine. The demo was great. But while the project team was watching adoption charts, the resistance was organizing, and it just walked into the room with a quorum. This lesson is about making sure you are never the person holding that slide, because the ambush was mappable in week one, and the map would have cost you two hours.

Resistance Is Information, Not Insubordination

Start by throwing out the frame most project teams carry into an AI rollout: that resistance is a character flaw in your colleagues, a stubbornness to be overcome, a communications problem to be messaged away. That frame is not just unkind. It is analytically useless, and it will get your pilot killed.

Resistance is information about your project. When a fifteen-year accounts payable (AP, the finance team that processes and pays supplier invoices) veteran goes quiet in meetings about the automation pilot, she is not being difficult. She is telling you, in the only channel available to her, that the pilot as currently designed proposes to take something from her, and nobody has told her what she gets back. That is a design defect in your project, surfaced early and for free. Treating it as an attitude problem is like treating a smoke alarm as a noise complaint.

And resistance obeys a rule every operations professional already knows from other domains: the earlier you find it, the cheaper it is. You know this rule from quality management, where a defect caught at the workstation costs a rework minute, the same defect caught at final inspection costs a batch, and the same defect caught by the customer costs the account. Resistance follows the identical cost curve. Found in week one, it is a design input: you adjust the rollout, change who is in the room, redesign a verification step, and it costs you a conversation. Found in month five, it has compounded into coalitions, hardened positions, and a steering-committee ambush, and it costs you the pilot and a piece of your reputation.

Here is the part that separates operators from optimists: unmapped resistance does not disappear. It organizes. It organizes into three recognizable shapes, and you have seen all of them.

• Quiet non-adoption. Nobody objects. Nobody complains. People simply do not use the thing, or use it just enough to appear compliant. This is the texture of MIT's famous finding that 95 percent of enterprise generative AI pilots deliver no measurable return: in a large share of those failures, the tools worked and the people politely declined to change, a pattern MIT called adoption without transformation. Silence is not consent. Silence is the cheapest form of veto ever invented.
• The workaround economy. The official process runs through the new tool; the real work flows around it through side spreadsheets, direct messages, and "just this once" exceptions that become permanent. Six months in, you have two processes: the one on the slide and the one that ships, and the second one has more users.
• The organized ambush. The concerns that were never voiced to you get voiced to each other, then to a sponsor, then in a meeting where the agenda item after yours is your pilot's obituary. By the time resistance is coordinated enough to appear on a steering-committee agenda, you are not managing change anymore. You are negotiating a surrender.

The arithmetic behind all of this is the arithmetic of the whole program. BCG's 10-20-70 rule holds that AI success is 10 percent algorithms, 20 percent technology and data, and 70 percent people and process. Resistance lives squarely inside the 70. A team that spends weeks stress-testing its model and zero hours mapping who stands to lose what has budgeted for 30 percent of its actual problem.

The Stakeholder Heat Map: Empathy Tooling, Not Enemy Intelligence

This lesson's artifact is the Stakeholder Heat Map: one row per stakeholder (or, better, per role, and we will come back to why), five columns that force you to know what you actually know, and a discipline for handling the result. It makes resistance visible in week one, while it is still a design input instead of a siege.

Before a single cell gets filled in, you need the framing that governs the entire artifact, because a stakeholder map is the most dangerous document this program will ask you to create, and it is dangerous in direct proportion to how badly you frame it.

The map is empathy tooling, not enemy intelligence. Its purpose is to understand what each person stands to lose from your project, in their terms, so that you can design the project so they lose less. That is the whole purpose. It is not a targeting system. It is not a list of obstacles. The moment a stakeholder map starts labeling people as problems, it stops being analysis and becomes a liability with your name in the file properties. A leaked map that reads like an enemies list ends careers, and it ends the author's career first, because everyone named on it will remember exactly one thing about you forever, and it will not be your process-mapping skills.

So three handling rules apply before any content rules, and they are not optional:

• Treat the map like HR data (HR: human resources, the function that holds personnel records), because functionally that is what it is: a document containing judgments about named individuals. That means restricted circulation, a named short list of who may see it, no posting to the shared project channel, no pasting into slide decks, and no leaving it open during screen shares. If your organization has rules for handling performance data, borrow them wholesale.
• Map roles, not personalities, wherever possible. "AP team lead" and "Treasury director" carry almost all the analytical value of the individuals' names with a fraction of the blast radius if the document escapes. You will sometimes need names for people whose influence is purely personal, but names should be the exception you justify, not the default.
• Every judgment carries its evidence. A disposition rating with no evidence line is gossip in a spreadsheet. A rating that cites an observable quote, behavior, or decision is analysis. This rule turns out to be both the ethical safeguard and the quality control, which is the kind of coincidence you should learn to trust.

Map roles, not enemies. Rate on evidence, not vibes. And give every red cell a move, because a map without moves is just surveillance.

With the framing set, here are the five columns. Each one earns its own slow treatment, because each one corrects a specific, expensive default error.

Reading the Five Columns

Influence: The Org Chart Lies

The first column asks: whose opinion moves other people's behavior? Not who outranks whom. The org chart records reporting lines; influence flows through trust lines, and the two maps overlap far less than leadership likes to believe.

The diagnostic question that finds real influence is this: who do people actually check with before changing how they work? In every operational team there are one or two people whose informal sign-off is the real go-live gate. Sometimes it is a supervisor. Just as often it is the veteran two desks over who has seen four systems come and go, or the person who fixes everything when it breaks, or the ex-supervisor who stepped down but never stopped being consulted. When that person shrugs and says "it's fine, we tried something like this in 2019, it didn't stick," the pilot loses twelve users in one coffee break, and your comms plan never even gets a vote.

Rate influence high, medium, or low, and rate it against the specific population your pilot touches. A Chief Financial Officer has enormous influence over budget decisions and approximately none over how a clerk actually keys an invoice at 4 p.m. on a Friday.

Disposition: Evidenced, Never Guessed

The second column places each stakeholder on a five-point scale: champion, supportive, neutral, skeptical, resistant. And here is the rule that separates a professional map from a Ouija board: every rating must carry an evidence line, a specific quote, behavior, or decision that justifies it. Not "seems negative in meetings." Not "I get a bad vibe." An observable fact a colleague could check.

This is where the previous lesson pays its dividend. If you ran an interview synthesis on your pilot's affected teams, you are sitting on a pack of themes with counts and quotes: "blame fear, raised by 11 of 14 speakers," with the exact language people used. That is disposition evidence of the highest grade, and AI is genuinely excellent at the first-pass conversion. The prompt is simple: from this synthesis pack, propose a disposition rating for each stakeholder or role, and cite the exact quotes that justify each rating; where the evidence is thin or absent, say so explicitly rather than guessing. Ten minutes later you have a drafted disposition column with citations.

Then comes the human's actual job: challenge every rating. Click through to every cited quote and ask whether it really supports the rating or just decorates it. Because there is a trap here, and it is precisely the false-confidence trap this program keeps warning you about: an AI will happily infer disposition from tone, and tone is terrible evidence. A person who asks hard questions in a workshop might be your most engaged future champion doing due diligence; the model reads "challenged the timeline twice" and drafts "skeptical." A person who says supportive things in every meeting might be managing you; the model reads the transcript and drafts "champion." The AI drafts. The human, who knows that the workshop skeptic later volunteered for the working group, corrects. A disposition column you did not challenge line by line is not a map. It is a horoscope with a confidence interval.

One more honesty rule: a blank cell is a legitimate rating. If you have no evidence about someone's disposition, the professional entry is "unknown," which is also a to-do item: go have the conversation. Filling every cell to make the map look finished is exactly how maps become fiction.

Blocking Power: Who Can Actually Stop This

The third column asks the bluntest question on the map: can this person stop or stall the pilot, regardless of where they sit on the org chart? Blocking power and rank are different currencies, and confusing them is the single most common stakeholder-analysis error, because project teams instinctively manage upward while the actual veto sits sideways or below.

The AP veteran who trains every new clerk has more blocking power over an invoice automation pilot than two directors combined. The directors can express concern in meetings. She controls what forty future employees learn on day one, which exceptions get worked first, and whether the new tool gets described to newcomers as "the system" or as "that thing corporate makes us log into." No slide can outrun that.

Blocking power comes in four operator-recognizable forms, and your map should note which form applies:

• Approval authority: a formal signature the pilot needs (security review, budget release, compliance sign-off). Visible, procedural, and at least it announces itself.
• Expertise monopoly: the only person who understands the legacy system, the exception rules, or the customer quirks the pilot must absorb. If they decline to share, the project starves politely.
• Informal veto: no formal authority at all, but the team takes its cues from them. The trainer, the veteran, the fixer. This is the form org-chart-driven teams miss every single time.
• Representative seats: union or works-council positions where applicable, which carry formal consultation rights over changes to how people work. In much of Europe, skipping this row is not a mapping error, it is a legal one.

Stake: What They Lose, in Their Terms

The fourth column is the heart of the artifact and the reason it qualifies as empathy tooling. For each stakeholder, write down what they stand to gain or lose from the pilot, in their terms, not yours. Your terms are cycle time and error rates. Their terms are time, status, craft, control, and job security, and the last four never appear on your dashboard.

Do this seriously and something uncomfortable and useful happens: most resistance turns out to be rational. Take the AP veteran again. The pilot proposes to automate exception triage, which is precisely the part of her job that earns her respect. Anyone can key an invoice; she is the one who can look at a mangled three-way mismatch and know in eight seconds which of five failure patterns produced it. That judgment is her craft, her status, and her Tuesday. Your pilot, as pitched, proposes to relocate it into a model. Her resistance is not a flaw in her character. It is an accurate reading of the design, and frankly a better reading than the project team's, because she noticed the problem first.

Naming the stake honestly is what makes a design answer possible. The answer to "the pilot deletes my craft" is not a pep talk about embracing change. It is a redesign: she becomes the verification authority, the human gate the AI's triage must pass, the author of the exception taxonomy the model is trained to follow. Her craft is not deleted; it is promoted from doing the sorting to defining and auditing the standard the sorting must meet. You met this idea in the Level 1 people-readiness lesson, and Level 4's change-management method builds a whole discipline on it. The heat map is where it becomes cell-by-cell concrete: one stake, one design answer, per row.

The Move: What Separates a Map From Gossip

The fifth column is the one that makes the other four defensible. For every cell that worries you, and without exception for every red cell, write the move: the specific design choice or conversation that addresses that stakeholder's stake. Moves come in four kinds:

• Involve: give them a designing hand in the thing they fear. Co-design sessions, ownership of a standard, a seat at the verification gate. The strongest move, and the most underused.
• Redesign: change the pilot itself so the loss they anticipate does not happen. The moves column is where McKinsey's core finding, that fundamentally redesigning workflows is among the strongest drivers of AI impact, gets applied to people instead of process steps. Redesigning the workflow around someone's stake is not a concession to the project. According to the best evidence we have, it is the project.
• Inform: for stakeholders whose concern is uncertainty rather than loss, a direct, honest, specific conversation, early, before the rumor gets there first.
• Escalate: for stakes only leadership can address, take the question up honestly rather than managing around it. We will see exactly this case in the worked example.

Here is why this column is ethically load-bearing and not just tactically useful. A map that rates people without committing to moves is surveillance: you have catalogued your colleagues' fears and done nothing with the catalogue except watch them. A map where every red cell has a move is a design document: you have found where your project causes loss and committed to changing the project or having the honest conversation. Same rows, same columns, opposite artifact. If you ever find yourself unable to write a move for a red cell, that cell is telling you something important about either the pilot or the limits of your authority, and both messages beat silence.

Where AI Helps, and Where It Must Never Go

AI earns its seat in this exercise in three specific places, all of them drafting and pattern work, none of them judgment.

• First-pass ratings with citations. As above: feed the synthesis pack, get proposed dispositions with exact supporting quotes, and challenge every one. The citation requirement is non-negotiable, because it converts the AI from an oracle into a research assistant whose work you can check.
• Diffing dispositions over time. A heat map is a snapshot, and stakeholder weather changes with every all-hands, org announcement, and pilot demo. Re-run the exercise after each significant project communication and ask the AI to diff the versions: which stakeholders' evidence has shifted since the last map, in which direction, and citing what? A skeptic warming up two weeks after the job-security town hall is a signal your comms worked. A supporter cooling off after the vendor demo is a signal worth a coffee this week, not a postmortem next quarter.
• War-gaming. Give the model the finished map and ask the operator's question: given this map, where does the rollout stall first? The model is genuinely good at tracing failure chains a tired project lead skips past: "if the AP team lead is skeptical and she trains all new clerks, your month-three onboarding cohort arrives pre-loaded with her skepticism; the workaround economy starts there." You are not obliged to believe the war game. You are obliged to check the two or three chains it surfaces.

And now the hard boundary, stated as bluntly as the brief demands: never feed personnel records, performance ratings, disciplinary history, compensation data, or anything HR-restricted into this exercise. Not into the map, not into the prompt, not "just for context." You learned a privacy pre-flight in the data chapter of this level; for people data it applies double. The map runs entirely on observable workplace evidence: what people said in interviews they consented to, what they did in meetings, what decisions they made. The moment HR-grade data enters an AI prompt, you have created a data-protection incident with a timestamp, and in works-council jurisdictions quite possibly a legal one. There is no analytical payoff on the other side of that line that survives the cost of crossing it.

Verification, as always in this program, is a two-step ritual, not a feeling. Step one: every rating traces to observable evidence, a quote, a behavior, a decision, and any cell that cannot produce its evidence gets blanked, not defended. Step two: the map gets reviewed with a second person who actually knows the politics before anything is acted on. You are one observer with one vantage point, and some of your ratings are wrong in ways only someone with a different vantage point can see. The reviewer who says "you have the Treasury director as your biggest threat, but she defers completely to her deputy on process matters, your map is missing the deputy" has just saved you a quarter.

Two Hours, Eleven Stakeholders, One Real Blocker: A Worked Example

Here is the whole method run end to end, with illustrative numbers, on the invoice-exception pilot this level has been following. Treat every figure as hypothetical; the pattern is the payload.

The transformation lead has the synthesis pack from fourteen interviews. On a Tuesday afternoon she builds the map: 45 minutes prompting the AI for first-pass dispositions with citations, 40 minutes challenging every rating against the quotes, 35 minutes on blocking power and stakes with the process map open beside her. Call it two hours for eleven stakeholders, mapped by role. A condensed excerpt of what the map showed:

The map's headline finding contradicted the project team's instinct, which is what maps are for. The team had been managing the Treasury director, because she was the most senior skeptic and the loudest voice in steering. The map said: high influence, yes, but low blocking power over this specific process, and her stake (control and audit exposure) is addressable with a design walkthrough of the verification gate. She is a conversation, not a crisis.

The real blocker was the AP team lead: moderate influence on paper, decisive blocking power in fact, because she owns the queue assignments that decide which exceptions the pilot even sees, and she trains every new clerk the department hires. Her stake read: craft (the exception triage judgment that makes her the person everyone asks) and blame exposure (when the model mis-sorts and a payment goes out wrong, she believes, correctly, that the finger will find her). The synthesis pack backed it: she was one of the eleven of fourteen speakers behind the blame-fear theme.

The move was involve, at full strength: she was asked to co-design the exception categories the model would be trained to sort into, and to own the verification gate, with mis-sorts logged as feedback she controls rather than errors pinned on her. Six weeks later (illustrative, but this arc is the standard one), the map's reddest cell had become the pilot's loudest advocate, demonstrating the queue to visiting managers from another site. Nothing about her character changed. The design changed. Her stake flipped from loss to ownership, and her disposition followed it, because it was rational both times.

And then the contrast cell, which matters just as much, because a lesson that implies every resistor can be flipped is selling you a manipulation kit. The senior AP clerk's stake was headcount fear grounded in a real event: an earlier automation was followed by a real layoff in 2023, and no comms plan on earth can message away a thing that actually happened. There is no clever move here, and pretending otherwise is how trust dies. The honest move is escalate: take the question to the sponsor and get a straight answer about jobs, whatever that answer is, and deliver it plainly. If the answer is "no reductions, redeployment to the verification and vendor-quality work, in writing," the fear has a real response. If the sponsor will not say it, then the clerk's resistance is not a stakeholder-management problem. It is the correct reading of an organization that has not decided to deserve the pilot, and better to learn that in week one than in month five. A map is not a manipulation kit. Some stakes only leadership can address, and the map's job is to route them to leadership fast.

The Map That Leaked: A Failure Story

Now the other path, assembled from real patterns and told with the details filed off, because versions of this story circulate in every large organization that has attempted a transformation.

A transformation lead at a mid-size services firm builds a stakeholder map for an AI rollout. He builds it fast, alone, and private, which feels prudent and is actually the first mistake, because privacy without discipline just means nobody challenges your ratings. The spreadsheet has a column literally headed "Blockers." It has full names. It has red cells. It has no evidence lines, because he knows these people, why would he need citations. It has no moves column, because the map's implicit purpose is not design. It is a list of who is with him and who is against him.

It leaks the way these things always leak: not through malice but through a thumbnail. A screen share in a routine call, the wrong monitor, four seconds of the file's tab visible, one attendee who recognizes her own name next to a red cell. That is all it takes. By Friday everyone named on the sheet has heard about it, and each has heard the version where their own cell is the reddest.

The pilot dies of the map. Not of the technology, which was never even discussed again. In the next steering meeting, three department heads who had been neutral, and one who had been quietly supportive, arrive aligned and cold, because whatever their views on invoice automation, they have a very clear view on being catalogued as obstacles by a colleague. Three months of relationship-building evaporates in one meeting. The lead is reassigned within the quarter, and the word "stakeholder mapping" becomes radioactive at that firm for two years, which is the cruelest cost of all, because the next project needed a map and could not build one.

Run the postmortem carefully, because the lesson is not "don't map." An unmapped rollout at that firm would likely have died anyway, just slower and with more ambiguity about why. The lesson is the discipline this chapter has been teaching all along: map roles where possible, attach evidence to every rating, give every red cell a move, and handle the artifact with HR-grade care. Notice that the discipline and the ethics are the same list. A map built on evidence and moves reads, if leaked, like what it is: a project team taking its colleagues' concerns seriously enough to redesign around them. There is no "Blockers" column to be ashamed of because there is no "Blockers" column. The safest stakeholder map to be caught holding is one you would be willing to explain to any person named on it, and that is not a coincidence. It is the test you should apply before you save the file.

One forward pointer before the Monday list, because it reframes several of your red cells for free: some of what reads as resistance is not about loss of status or control at all. It is skill fear, the quiet dread of being made to look incompetent in front of a tool everyone else seems to get. Skill fear trains away, which makes it the cheapest red cell on the whole map, and the next lesson, on AI-assisted skills-gap analysis, is how you find it and cost it.

What to Do Monday Morning

Two hours, one artifact, this week, for whatever pilot or proposal you are closest to.

1. List your pilot's stakeholders, aiming for 8 to 12 rows. Use roles, not names, wherever the role carries the information. Include the informal figures the org chart hides: ask yourself who people actually check with before changing how they work, and make sure that person has a row.
2. Draft dispositions only where you hold evidence, and leave the rest blank. If you have interview themes or a synthesis pack, let AI propose first-pass ratings with exact supporting quotes, then challenge every single one. A blank cell is honesty; a guessed cell is a horoscope. Each blank is also your interview to-do list.
3. Identify the two highest blocking-power cells on the map. Ignore rank while you do it. Ask who can stall this pilot through approval authority, expertise monopoly, informal veto, or a representative seat. These two rows are where your next two weeks of relationship work belong.
4. Write one move per red cell: involve, redesign, inform, or escalate. If a red cell resists every move you can think of, that is a finding about the pilot or about what only your sponsor can answer. Route it upward honestly instead of managing around it.
5. Review the map with one politically fluent colleague before acting on any of it, under the same confidentiality you would give performance data. Ask them two questions: which rating is wrong, and who is missing. Then, and only then, start making moves.

Key Takeaways

• Treat resistance as information about your project, not a character flaw in your colleagues: it follows the quality-cost curve, cheap to address in week one and ruinous by month five, because unmapped resistance does not disappear, it organizes into quiet non-adoption, workaround economies, and steering-committee ambushes.
• Build the Stakeholder Heat Map with five columns per row: influence, evidenced disposition, blocking power, stake in the person's own terms, and the move that addresses it, and remember BCG's 10-20-70 rule puts this work inside the 70 percent of AI success that is people and process.
• Rate influence by asking who people actually check with before changing how they work, because the org chart records reporting lines while adoption follows trust lines, and the two rarely match.
• Anchor every disposition rating to observable evidence, a quote, a behavior, or a decision, let AI draft first-pass ratings with citations from your synthesis pack, and challenge every one, because AI-inferred disposition from tone is the false-confidence trap in miniature; leave cells blank where evidence is missing.
• Measure blocking power separately from rank: approval authority, expertise monopoly, informal veto, and union or works-council seats can each stall a pilot, and the AP veteran who trains every clerk outweighs two directors.
• Write a move for every red cell (involve, redesign, inform, or escalate), because the moves column is McKinsey's workflow-redesign finding applied to people, and a map without moves is surveillance rather than design.
• Keep HR-restricted data out of the exercise entirely, never feed personnel records or performance data into any prompt, circulate the map like HR data, and prefer roles to names, so that the map, even leaked, reads as empathy rather than an enemies list.
• Verify before acting: trace every rating to its evidence, review the map with a second person who knows the politics, and escalate the stakes only leadership can answer, because a heat map is a design instrument, not a manipulation kit.