Chapter 4
Where does power actually come from?
On Friday, November 17, 2023, Sam Altman joined a video call expecting a routine board meeting and was fired from OpenAI — the company he ran, at the moment it was the most talked-about company on Earth. The board's authority was unquestionable: it held every formal right to remove the CEO, and it used them, announcing that Altman had not been "consistently candid" in his communications.1 What followed was a live demonstration of the difference between authority and power. Co-founder Greg Brockman quit within hours. Over the weekend, employees flooded X with a single coordinated sentence — "OpenAI is nothing without its people." On Monday, Microsoft announced it would hire Altman and Brockman to lead a new AI unit, and would take any OpenAI employee who wanted to follow at matching pay.13 By Monday more than 700 of OpenAI's roughly 770 employees — including executives, and eventually even the co-founder who had helped orchestrate the firing — signed a letter telling the board: resign and reinstate Altman, or we all leave. "Your actions have made it obvious that you are incapable of overseeing OpenAI," it read.3 Late Tuesday night, five days after the firing, Altman was CEO again. The board that fired him was gone.2
Two years earlier, in December 2021, baristas at a Starbucks store in Buffalo, New York, won the first union election in the company's modern history — the start of the most successful new organizing campaign in recent American labor.5 Four years later, the movement had grown to 667 unionized stores and more than 9,500 baristas.68 It had also, as of that fourth anniversary, produced zero signed contracts. Baristas struck before Christmas 2024 across hundreds of stores, struck again over a dress code in May 2025, and on November 13, 2025 launched the longest strike in company history — 131 days — before bargaining was set to resume in April 2026.569 The workers had done everything the OpenAI employees did: organized, coordinated, threatened exit, walked out. They had legal certification the OpenAI employees never needed. And they were still waiting.
Put the two cases side by side and the puzzle sharpens into a paradox. A board with total formal authority lost everything in five days to employees with no legal standing at all. Employees with formal legal standing — federally certified union representation — could not convert it into a contract in four years. If power were the org chart, the OpenAI board should have won. If power were collective action, the baristas should have won quickly. Something else is doing the work, and reasonable observers disagree about what: talent scarcity, outside options, timing, public sympathy, or simple corporate stubbornness. This chapter gives you the concepts to replace those guesses with analysis.
The central question: where does power in organizations actually come from, and what determines whether it works? The tools: the classic bases of power; dependence theory — the single most important idea in this chapter; influence tactics and political skill, the behavioral craft of getting your way; organizational politics and what the evidence says it does to people; what holding power does to the powerful; and countervailing power — how the powerless combine. Then we return to OpenAI and Starbucks, two collective revolts with opposite speeds, and let dependence theory explain the difference.
Power is the capacity to influence others' behavior even against resistance; authority is the subset of power granted by formal position. The founding taxonomy comes from John French and Bertram Raven's 1959 essay, still the field's common vocabulary.10
• Legitimate power — from position: the right to direct, rooted in others' acceptance of the structure.
• Reward power — control over things people want: pay, assignments, recognition.
• Coercive power — control over punishments: discipline, termination, exclusion.
• Expert power — from knowledge and skill others need.
• Referent power — from identification, admiration, and loyalty: people comply because of who you are to them.
The taxonomy's enduring lesson is that the five bases behave differently: position-based power (legitimate, reward, coercive) produces compliance that lasts exactly as long as surveillance and the position do, while person-based power (expert, referent) produces commitment that travels with the person. The OpenAI hook is the taxonomy made flesh: the board held legitimate, reward, and coercive power in full; Altman, stripped of position, retained expert and referent power — and the employees' loyalty followed the person out the door, not the position that remained.13 The framework's limit is that it names the currencies without explaining the exchange rates: it cannot tell you why referent power beat legitimate power at OpenAI but coercive power has held off referent-rich organizing at Starbucks for four years. For that you need dependence.
Richard Emerson's 1962 power-dependence theory states the field's most important equation: the power of A over B equals the dependence of B on A.11 Power is not a property of a person; it is a property of a relationship, determined by two things: how much B values what A controls, and how many alternatives B has for getting it. Pfeffer and Salancik's resource dependence theory scaled the same logic to whole organizations: firms are controlled by whoever supplies the resources they cannot do without.12
• Criticality — how essential is the resource you control to the other party's goals?
• Scarcity / substitutability — can they get it elsewhere, or replace it? Irreplaceability is power.
• Discretion — do you actually control the resource's allocation?
• Balancing operations — Emerson's routes for the weaker party: reduce your dependence (want it less, find alternatives), or raise theirs (become more valuable, or form coalitions so you must be dealt with as one).11
Dependence theory answers questions the org chart cannot: why a technician who alone understands the legacy system outranks his title; why departments that manage the firm's most critical uncertainties collect outsized budgets; why your negotiating position in any job depends less on your performance than on your outside options. It converts power analysis into two questions you can ask about anyone: what do they control, and who else could supply it? Its limits: dependence explains the capacity for power, not the will or skill to use it (that is Section 2.3), and treating all relationships as resource exchanges misses the ways legitimacy and shared identity make power feel like something other than a transaction — which is precisely what lets it operate quietly.
Power is potential; influence tactics are behavior. Research led by Gary Yukl catalogued the recurring tactics managers use on bosses, peers, and subordinates and — crucially — measured which ones work.15
Three tactics consistently outperform the rest: rational persuasion (evidence and logic), inspirational appeals (linking the request to values and aspirations), and consultation (inviting the target to help shape the plan). The weakest are the ones novices reach for first: pressure, coalition used bluntly, and naked legitimating ("because I'm in charge").15 Soft, target-engaging tactics generate commitment; hard tactics generate, at best, compliance and, routinely, resistance. Effectiveness also depends on a person-level capacity that Gerald Ferris and colleagues termed political skill: social astuteness, interpersonal influence, networking ability, and — the load-bearing dimension — apparent sincerity, since influence attempts that read as manipulation fail regardless of technique.16
This literature converts "office politics" from a personality trait into a learnable craft — and its findings are mildly comforting: the tactics that work best are the most respectful of the target. The cautions: most tactic studies rely on self- and observer reports rather than experiments, so the causal arrow between tactic use and success is partly inferred; and political skill's dark edge is real — the same capacities that make someone an effective advocate make an effective manipulator, which is why Section 2.4's question (what does a political climate do to everyone else?) matters.
Organizational politics is the use of power to influence decisions through means outside formal roles — lobbying, coalition-building, information control, impression management. Researchers study it largely through perceptions of organizational politics (POPs): the degree to which employees see their workplace as governed by self-interest and connections rather than merit.
The meta-analytic verdict, assembled by Chang, Rosen, and Levy across dozens of studies, is consistently negative for the perceivers: the more political employees judge their organization to be, the lower their job satisfaction and organizational commitment, the higher their strain, and the stronger their intentions to quit — effects that flow through a sense that the environment is unpredictable and effort–reward links are broken.17 Note the mechanism: perceived politics is toxic largely because it destroys the instrumentality that Week 3's expectancy theory identified as motivation's load-bearing wall.
Here the chapter must hold two truths at once. Perceived politics damages the workforce — and politics is also how decisions actually get made wherever goals conflict and resources are scarce, which is everywhere. Jeffrey Pfeffer's long-standing argument deserves its place: managers who refuse to build coalitions, court allies, and manage impressions do not thereby make their organizations meritocratic; they merely lose to those who do, and their good ideas die in committee.18 The reconciliation is about visibility and fairness: politics experienced as a rigged game corrodes; influence exercised through consultation and persuasion, inside procedures people consider just (Week 3's procedural justice), reads as leadership. The same behavior can be either, depending on how it is done and who is watching.
Power changes the person who holds it. Dacher Keltner, Deborah Gruenfeld, and Cameron Anderson's approach/inhibition theory synthesizes a large experimental literature: elevated power activates approach tendencies — action orientation, optimism, disinhibition — while low power activates inhibition, vigilance, and constraint.14 Downstream, experiments find the powerful attend less carefully to others, stereotype more, take others' perspectives less accurately, and overestimate the support they enjoy. The finding has a mirror image that Rosabeth Moss Kanter argued a generation earlier: powerlessness corrupts too. Managers who lack real capacity — resources, information, backing — become controlling, territorial, and rule-obsessed, because petty domination is the only power left to them.13
Together these findings explain a pattern you will see repeatedly: boards and executives blindsided by revolts they never saw coming (approach-oriented confidence plus degraded perspective-taking is precisely the recipe for firing a beloved CEO and being astonished when 95 percent of the company rebels), and micromanaging supervisors created not by bad character but by hollow positions. The corrective is structural rather than moral: audit the powerful's information diet (who can tell them no?) and give the "powerless" managers real resources rather than exhortations. Caveats: most of this evidence is laboratory-based with situationally primed power, and effects vary with the holder's values and accountability — power amplifies the person, it does not replace them.
Emerson's final balancing operation — coalition formation — is the theory of every strike, union, and mass resignation threat: individuals who are each substitutable become collectively critical when they act as one.11 But the dependence equation disciplines the romance. A coalition's power still equals the target's dependence on it, which means three variables decide outcomes: coverage (what fraction of the critical resource does the coalition control?), substitutability (can the target replace the coalition's resource — hire replacements, automate, wait?), and staying power (can members afford to withhold longer than the target can afford to lose them?). Labor law adds a fourth: legal certification obliges an employer to bargain, but in the American system it cannot compel agreement — a gap through which years can pass. Collective action, in short, is not a separate kind of power; it is dependence engineering, and it succeeds exactly insofar as it makes the arithmetic bind.
These cases were chosen because they hold the tactic constant and vary the structure. Both are employee collectives confronting their organization's formal power-holders with a credible mass-exit threat. If power flowed from formal authority, the OpenAI board should have prevailed and Starbucks should have been forced to sign long ago (the baristas hold federal certification; the OpenAI staff held nothing). If power flowed from collective courage, both should have won. Dependence theory predicts what actually happened — and the comparison shows why it, not the org chart, is the master key.
The facts: fired November 17, 2023, Altman was back by November 22 with a reconstituted board.12 The mechanism ran straight through Emerson's equation. OpenAI's entire value was its people — a few hundred researchers who could not be replaced at any price on any timeline; the employees' letter said so in one sentence the whole company had already tweeted: "OpenAI is nothing without its people."13 Microsoft's offer to hire Altman, Brockman, and every defector at matching compensation did the decisive structural work: it gave the coalition a costless outside option, driving the employees' dependence on OpenAI toward zero while the board's dependence on them remained absolute.13 Coverage: over 700 of roughly 770 — about 95 percent, including the executives running the company and, remarkably, the co-founder who had led the firing.23 Substitutability: none. Staying power: irrelevant — the exit was instantaneous and individually profitable. The board's legitimate, reward, and coercive power evaporated because every one of those bases presupposes the other side needs something you control. By Tuesday night it controlled nothing anyone needed.2
The facts: from the first Buffalo victory in December 2021, the campaign grew to 667 unionized stores by the end of 2025 — a historic organizing achievement, yet about 6 percent of Starbucks's more than 10,000 company-owned U.S. stores, and roughly 9,500 baristas of a workforce over 200,000.68 A much-heralded February 2024 framework agreement promised a path to contracts; talks stalled.9 The union escalated: a pre-Christmas strike across hundreds of stores in December 2024, a 50-plus-store dress-code strike in May 2025, and the open-ended "Red Cup Rebellion" beginning November 13, 2025 — 131 days, the longest strike in company history — before the parties set an April 2026 return to the table, with the union having delivered a full proposed contract ($17 wage floor, 4 percent annual raises, grievance procedures).5679 Four years after Buffalo: still no first contract — a fate the campaign shares with other celebrated recent wins, including Amazon's Staten Island warehouse.5
As a theory test: every dependence variable that favored the OpenAI coalition runs the other way here. Coverage: 6 percent of stores cannot threaten the enterprise; 95 percent of OpenAI could end it. Substitutability: barista labor, however skilled and however unjustly treated, is replaceable in weeks in most labor markets; frontier AI researchers were replaceable never. Outside options: no Microsoft stands ready to hire 9,500 baristas at matching pay; individual exit means another service job, which is why exit here takes the costly form of striking rather than quitting. Staying power: hourly workers forgo wages every strike day, while a company with more than 10,000 U.S. stores absorbs temporary closures it publicly called immaterial.7 The union's real countervailing assets are the ones dependence theory locates outside the employment exchange: legal process (certification compels bargaining, unfair-labor-practice charges impose cost), reputation (public sympathy, boycott calls, shareholder pressure on directors7), and time — expanding coverage store by store. That is why the baristas' war is long where the researchers' war was a weekend: not weaker courage, weaker dependence.
| OpenAI staff (Nov 2023) | Starbucks Workers United (2021– ) | |
| Formal standing | None — no union, no legal claim | Federal certification; employer legally must bargain |
| Coverage of critical resource | ~95% of an irreplaceable workforce | ~6% of stores; ~5% of workforce |
| Substitutability | Effectively zero | High in most local labor markets |
| Outside option | Microsoft: matched pay for all, immediately | Comparable service jobs; exit gains little |
| Result | Board replaced, CEO restored in 5 days | 667 stores organized; longest strike in company history; no contract after 4 years |
| Power source that decided it | Dependence (expert scarcity + coalition + exit option) | Contest still running on law, reputation, and coverage growth |
Three disciplined conclusions. First, the pairing vindicates Emerson over the org chart: identical tactics produced opposite results because the dependence structures differed, which is exactly what a relational theory of power predicts and a positional one cannot explain. Second, beware the halo on the OpenAI story: five-day victories belong to workers who are scarce, and celebrating the researchers' leverage without noticing its source quietly blames less-leveraged workers for lacking what the labor market never gave them. Third, the Starbucks case is unfinished, and honest analysis says so: growing coverage, accumulating legal findings, reputational pressure, and a delivered contract proposal are all dependence-raising operations whose sum may yet bind — or may not. The theory tells you which variables to watch; it does not guarantee the ending.
Return to the central question: where does organizational power come from, and what determines whether it works? Four conclusions travel.
Map dependence before you read titles. For any actor, ask what they control, how critical it is, and who else could supply it.11 Every surprising power outcome in this chapter — a fired CEO stronger than his board, certified workers waiting four years — is unsurprising once the dependence map is drawn.
Build person-based power; treat position-based power as rented. Expert and referent power follow you; legitimate, reward, and coercive power belong to the chair, not its occupant, and vanish the moment others stop needing what the chair controls.10 For your own career this cashes out concretely: cultivate scarce skills, real relationships, and outside options — they are your dependence position.
Practice influence as a craft, in the open. The best-performing tactics — evidence, aspiration, consultation — are also the most legitimate ones, and political skill without apparent sincerity self-destructs.1516 Politics conducted inside fair procedures is leadership; the perception of a rigged game measurably corrodes everyone underneath it.17
Audit what power is doing to you and around you. Holding power degrades perspective-taking exactly when you need it most; lacking real power breeds the petty control everyone mistakes for bad character.1314 The managerial response to both is structural: keep people around you who can say no, and give your subordinates enough genuine power that they never have to hoard the fake kind.
Notes appear as superscript numbers in the text and correspond to the numbered sources above. DOIs are provided where available; classic books and chapters are cited to their original publishers.