Measured against the research on self-managing organisations

A working draft, November 2026; expect it to change. One of a series weighing what we've built against the work of people whose thinking we respect.

In this page

• What's being measured: an operating system, the networks of organisations running it, and The Grove, the first of them
• Our understanding of ten pieces of research on organisations without managers, which we'd welcome being corrected on
• Nine tests from that research in turn: what we chose, why, how we think it lines up, and where we might be wrong
• The findings that are hardest for us, and the places we're least sure of

Most of what's written about organisations without bosses is advocacy. Michael Lee, whose research runs through this page, says much of the practitioner rhetoric can be "hyperbolic, promising occupational nirvana".¹ A researcher in this field should arrive at a book like this one expecting more of the same: a design by people who already believe, checked against nothing.

So this page is different from the others in the series. Those weigh the design against a body's ideas. This one weighs it against evidence. That includes a twelve-month field experiment that found no average benefit to people from taking decision authority away from their managers, a public service where conflict between self-managing teams stalled the work, and a firm whose chief executive brought managers back after three years.

We don't try to argue any of that away. We set out what we chose, why, and how we think it lines up, so that people who do this research can test our reasoning and tell us where it fails.

What's being measured

Three things, and it helps to keep them apart.

The operating system is the software an organisation runs: the rules for seats, decisions, trust, money and work, with a few guardrails at its core that nobody inside an organisation can vote away. It has no name of its own. We're building towards each organisation running its own copy; as the code stands, the simulated organisations share one space. Its guide inside every organisation is an AI companion called Wise-Eye.

A network is organisations running the operating system that recognise one another and share agreements. Each network has a Steward: an organisation that looks after the network's code and ships its releases.

The Grove is the first such network. It's still being built and tested, and its organisations today are simulated: seeded, grown and run by AI people. Nothing on this page has been tested with real organisations. That matters more here than anywhere else in the book, because this page is about evidence, and we have none of our own yet.

Three words used below. An organisation's work is held in branches: parts of the organisation shaped around a piece of work, each with a purpose and the things it answers for. Each branch has a seat, held by a steward. And decisions are made in forums. A branch is governed in one forum: the stewards of its parent and of its sibling branches, with the branch's own steward taking part as a person. The root has no parent, so it's governed by its own stewards and the stewards of the branches grown directly from it.

We keep to the rule the rest of the book follows: is only for what's built, building towards for what's decided but not built, and we don't know for what's open.

Our understanding of the research

This is how we understand ten pieces of work. If we've got any of it wrong, we'd like to be corrected, and we'll change the page.

The definition and the frame

• Lee and Edmondson (2017) reviewed decades of research on organising with less hierarchy and gave the field a working definition. A self-managing organisation is one of those that "radically decentralize authority in a formal and systematic way throughout the organization".²
• Martela (2019) takes four problems every organisation must solve, from Puranam, Alexy and Reitzig, splits two of them, and compares how bureaucracy, adhocracy and self-managing organisations solve all six: dividing tasks, allocating them, rewarding, stopping freeriding, giving direction and coordinating.³
• Reineke, Katila and Eisenhardt (2025) review a hundred and seventy-seven studies of decentralisation across four waves, and conclude that today's is "neither a revolution nor a mirage".⁴

Studies of single organisations

• Lee (2024) spent eighteen months with a payments company of a hundred and twenty-five people that had adopted Holacracy, asking which practices keep decentralised authority going and why they break down.⁵
• Martela and Nandram (2025) describe Buurtzorg, the Dutch home-care organisation with over fourteen thousand people in more than nine hundred teams, two directors and no middle managers.⁶
• Jensen and Mikkelsen (2024) studied a Danish public addiction rehabilitation service of about fifty professionals, which in 2018 chose not to replace two departing department managers and handed their work to its teams.⁷
• Valkiainen (2026) spent twelve months coaching and studying a Holacracy adoption in a family-owned manufacturer in the Baltic region.⁸
• Bourlier Bargues, Valiorgue and Islam (2025) followed a French dairy firm of eighty people for four years, through the removal of all its managers and their return.⁹

Comparisons and an experiment

• Lee (2019), in his Harvard doctoral dissertation, ran a twelve-month field experiment in the State of Washington's technology agency, asking whether taking decision authority away from managers improves people's sense of power, engagement and satisfaction, and for whom.¹⁰
• Lee and Young-Hyman (2026) compared four organisations committed to decentralising that had to centralise during the pandemic.¹¹

From these we've drawn nine tests:

1. The definition: authority decentralised radically, formally and throughout.
2. Dividing and allocating the work, with no manager to do either.
3. Rewards and freeriding, with no manager to judge or to watch.
4. Direction and coordination, and whether the system becomes the hidden boss.
5. Bounding and grounding authority, the practices Lee found working, and what they cost.
6. Conflict between the parts, which is where one organisation broke.
7. Who takes up authority, and who is harmed.
8. Crisis: centralising without losing the commitment.
9. What stays at the centre, moving across, and where it fits.

None of these authors certifies anything, so this isn't a compliance check. It's our own reading, and it may be a generous one. That's what the sections on where we might be wrong are for.

1. The definition

Our understanding. Three features make a self-managing organisation. The decentralisation is radical: the reporting relationship between manager and subordinate is gone, and nobody's decision rights can be overridden by someone because they're the boss. It's formal: written into explicit rules, because hierarchy "cannot easily be altered by simply declaring it absent", and authority that's merely delegated gets taken back in a conflict or a crisis.¹² And it runs throughout the organisation, senior people included.

Lee is careful about what this doesn't mean. Managerial work still has to be done by someone. Some roles still carry more authority than others. Informal hierarchy survives. What's gone is any formal relationship in which one person holds power over another's autonomy.

The Stanford review adds a warning from the other side. It places Holacracy as one variant of the "organic" form of organising, and finds that such forms have a right amount of structure. Firms that tried to do with very little became unstable, and it concludes that too little structure does more damage than too much. Among the firms it names is Zappos, which ran Holacracy, so formal rules alone didn't count as enough structure for these authors.¹³

What we chose, and why. There are no reporting relationships in the operating system at all. Nobody is anybody's manager, and nothing in it lets one person direct another's work.

Authority sits in seats and forums. Every change to an organisation's structure is a proposal, decided by consent in a forum, with the weight behind each voice given by peers and taken back by them. The founder holds no override: the root is one seat in a forum, like any other.

And the rules are as formal as rules can be. They're code, the same in every organisation, with guardrails no forum can vote away. We chose that because we'd seen what Lee describes: empowerment that depends on a manager's goodwill ends when the goodwill does.

How we think it lines up. We think it meets all three features, and carries more structure than the firms the Stanford review warns about, though that review counted a Holacracy adopter among them. In the cases Lee studied, the rules were a document people ratified. Here they're the system the organisation runs on, so there's no gap between the rules and the tool.

Where we might be wrong. Three places.

First, the rules only bite once a second person holds a seat. Until then one person decides the structure, the money and who is seated, and members can only advise. Nobody's work is being directed, because only people who hold seats carry work. But a founder who is slow to seat anyone keeps every decision, and the design shows that plainly without preventing it.

Second, branches form a tree, and the forum around a branch's parent can change what it answers for or fold it. That's a forum deciding, never a person, but it's still a hierarchy of parts, and someone who holds a branch above may step in when a promise below is at risk. A researcher could fairly ask how different that feels from a manager on a bad day.

Third, giving trust the status of weight in decisions may harden the informal hierarchy this research warns about, even though peers can take their trust back any day. We've designed trust to fade unless it's renewed, and that part isn't running yet.

2. Dividing and allocating the work

Our understanding. In a bureaucracy, managers break the goal into tasks and hand them down. In a self-managing organisation, Martela says, individuals have both the responsibility to spot what needs doing and the authority to make it part of the structure, and people have considerable say over which roles they take.¹⁴ That asks a lot. People have to be proactive, and committed enough that the unattractive but necessary work still gets done. Where work is tightly linked, something explicit is needed so that every necessary task ends up with someone.

The French dairy firm shows what happens when it doesn't. With managers' titles and most monitoring removed, output rose. But the former managers still faced clients and inspectors, now without the records and indicators they used to keep. Two outside audits found the gap, and it was the beginning of the end.⁹

What we chose, and why. Anyone can raise a tension, and a tension can become a proposal for a new branch, a changed purpose or a new thing a branch answers for. Branches are born from work that keeps arriving, with evidence, so the structure follows the work.

For allocation, work belongs to a branch, never to a person. Everything that arrives comes through one door and is routed to whichever branch has said it answers for that kind of work. Work that fits nowhere rises until it rests at the root, so nothing is left without a home.

Seats are offered, not assigned: an open seat is posted, members say they'd like it, someone else proposes them, and the forum decides. A step where the person named says yes isn't built yet. An empty seat is held by the branch above until someone steps in.

And a promise made to someone outside always belongs to a branch, with a day on it, and is carried by a named person or shows as needing one.

How we think it lines up. Closely on division, which is bottom-up by proposal as Martela describes.

On allocation we've gone a different way from his examples. People don't pick from one pool across the organisation, and they don't negotiate contracts with each other. The work finds the branch, and inside it whoever holds the seat picks it up. We think that covers his test that every necessary task ends up with someone, including the dull ones, because work a branch leaves or refuses rises to the branch above and never disappears.

And the dairy firm's gap is one the operating system is built to close. The work that passes through the system, the decisions and the promises are recorded as they happen. What happens outside it, on a production line say, is still only what people write down.

Where we might be wrong. Routing by what a branch has written down depends on the writing being good, and Lee's case shows people refusing anything not in their written role. We take that up in the fifth test.

Work with no home is meant to be a tension, not a burden. It shows where it rests, and a tension becomes a proposal: a new thing a branch answers for, or a new branch. Most of it should be given a home long before it reaches the root, and what does reach the root is there for the root to grow structure for, not to do. That's the design. What's untested is whether a busy root does the proposing or quietly does the work. Lee found this never settled in his case: the role that caught what was unassigned kept an edge nobody could draw, and one senior holder felt responsible for the whole company.¹⁵

And keeping records isn't the same as being able to answer an inspector. What the design offers is this. Each piece of work keeps its own thread, so its history is easy to see. Work can carry measures. And every proposal says what's expected to happen, as measures of the organisation's own, and is looked back on against them. Whether that adds up to what an auditor asks for is untested.

3. Rewards and freeriding

Our understanding. With no manager to set pay, Martela expects peers to do it, and he expects heavier reliance on work being rewarding in itself. With no manager to watch, he expects peers to monitor each other's performance, with explicit steps for resolving conflict and people skilled enough to use them. He's blunt about the stakes: without a way to deal with freeriding, cooperation declines.¹⁶

Buurtzorg does it differently from his prediction on pay, following the national agreement for its sector with no bonuses, but relies heavily on visibility. Every visit is recorded, each nurse can see what teammates have done, and a dashboard compares each team with the rest.⁶

The case studies show how hard the peer part is. In the Danish service, staff avoided giving each other negative feedback to protect solidarity, and once feedback between teams turned toxic they lost the ability to sanction poor work at all. In the dairy firm, workers still couldn't bring themselves to sanction a teammate and pushed their former managers to do it.¹⁷ And the older research carries a warning from the other direction: peer control in self-managed teams has been found to become as coercive as a manager's.¹⁸

What we chose, and why. Pay is a share of what the work earned, divided by a rule the organisation agrees by proposal in its root forum. The rule is a set of dials: how much goes by hours given, by roles held, by contribution, and in equal shares. Beneath that an organisation can set an equal floor for every member. Nobody's pay is set by a person above them. Trust and money are kept apart as two currencies, and neither converts into the other.

On freeriding we've chosen something the research might not expect. The operating system publishes no score on any person. There are no ratings, points, rankings or comparisons on show, and the organisation's reading of itself has no place for a person's name. What's held about a person's standing is the trust peers have given them in a forum, and a private picture of skill that decides only who is asked for advice. Neither is ranked, and the picture of skill is shown only to the person.

What's visible instead is the work, and everything a person touches in it: what each branch answers for, every promise and its day, who carries it, and what's waiting on each person, said as what waits and never as a judgement. A promise that slips rises as a message to the branch above, where someone may step in. And the consequences are real. Peers can take back the trust they've given someone, which takes away their weight in decisions. And a forum can hand a seat to someone else by proposal.

How we think it lines up. On rewards, partly. Nobody decides anyone's pay. The people in the root's forum agree one rule for everyone, by proposal, and each person's share then follows from the record: hours given and roles held. So peers set the rule, which is less than Martela's peers judging each other. The exception is the contribution dial, where an organisation turns it up.

On freeriding, closer than it first looks. Martela expects peers to monitor performance and to have a way to act. Here the work is open for peers to see, and peers hold both sanctions: trust and the seat. What we've refused is turning that into a published number, because of the warning this research itself carries about peer control, and because we think people hide what's scored.

Where we might be wrong. The case studies say the hard part isn't having a sanction. It's that peers won't use it. Trust withdrawn and a seat handed on both need someone to act against a colleague, and the Danish and French cases both show people who wouldn't.

A slipping promise rising by itself helps, because nobody has to accuse anyone for it to be seen. But what follows still needs a person.

A member who holds no seat has neither sanction to lose. The contribution dial, which is off unless an organisation turns it up, follows how peers see each person's contribution, and we haven't shown that this stays clear of the rating we've refused. A way to exclude someone who's acting in bad faith is something we're building towards; it isn't there yet.

And making everything a person touches visible is peer monitoring, whatever we call it. The research's warning about that applies to us too.

4. Direction, coordination and the hidden boss

Our understanding. If people are to make decisions a manager used to make, they need the information a manager used to hold. Martela's rule is that when decisions are decentralised, information must be too, finances included, and people need to be able to read it.

For coordination between teams, with no hierarchy to fall back on, he sees two answers: people given the job of linking, and information systems that track who's responsible for what and are open to everyone. Those systems, he says, have taken over much of the coordinating, progress-tracking and information-sharing that managers did.¹⁹

Buurtzorg bears him out. Its purpose-built system was designed with nurses, because existing ones served the administration and not the people doing the work.⁶ Lee credits technology too: in his case the platform holding the roles was visited a hundred thousand times in a year.²⁰

The Stanford review supplies the warning. On platforms, where an owner's software allocates the work, it finds what it calls an "invisible hierarchy". Tasks are handed out by algorithm, often with no explanation. The rules are a black box and keep changing. Monitoring is instant and impersonal, and information runs towards the owner.

Its verdict is that this kind of management may not decentralise power at all, and that it brings organising back towards bureaucracy.²¹ The review is describing marketplaces such as ride-hailing apps, not organisations like the ones here. But a system that routes work and applies rules automatically should expect to be read against that passage.

What we chose, and why. This is the centre of what an operating system is for. The organisation's purpose is held at the root and every branch has its own, so direction is written where everyone can see it. A founder's vision enters as bets, each saying what would show it wrong.

The organisation reads its own state from a ledger of work at each branch, open to everyone in it. Between branches, coordination happens through the routing of work, through conversations between branches, and through promises.

And Wise-Eye sits beside every seat, reading the same records the organisation keeps, routing what arrives, drafting replies and saying what's at risk. It advises and drafts, and where a person has granted it, it sends low-stakes routine work by itself. It never decides a proposal.

How we think it lines up. On Martela's test, this is where we're most confident. His claim that information systems can take over the manager's coordinating work is close to the premise of the whole design, and the transparency he asks for is the default.

On the review's warning, we think the design differs from a platform at each of the points it names. Work is routed by what each branch has itself written down that it answers for. Where the routing is a guess, it's marked as one, and nothing is sent while it's a guess. The branch where work lands can move it on or refuse it, and the routing learns from where people put things. Nobody is made to take work. What each branch answers for changes only by the organisation's own proposals; the order routing follows, and the guardrails, are in the code.

Nothing ranks or removes a person. And there's no owner on the other side of the screen: the organisation's reading of itself is open to everyone in it.

Where we might be wrong. One of the review's points lands squarely. In one of its examples it says the transparency of an algorithm is crucial, and the operating system's code isn't open source. We're building towards each organisation choosing whether to take each release; as the code stands, a release reaches all of them at once. Either way, an organisation can see what the system does but can't read how. That's a black box at the core, however open everything around it is.

Open information is also only useful to people who can read it, and Martela expects self-managing organisations to need more educated people for that reason. We're leaning on Wise-Eye to close that gap, and we don't know that it will.

And an AI that drafts and routes for every seat could become the manager nobody appointed. It can't decide a proposal, but what it chooses to show is a kind of power.

5. Bounding and grounding authority

Our understanding. This is Lee's most useful finding, and it runs against the usual picture. The standard idea of the organisation after bureaucracy is loose and informal: remove the hierarchy, blur the roles, let people adjust to each other. Lee found the opposite working. His conclusion is that "ambiguous and informal rules are often the foe, not the friend, of decentralization".²²

His published study treats decentralised authority as "a dynamic and situated achievement that must be continually enacted", and names what kept it going.²³ Roles were written down and published to the whole company, with every responsibility and its authority spelled out.

Then two things followed. People worked from those boundaries, which he calls bounding authority. Junior people could say no to senior ones by pointing at the written role, and senior people used the boundaries to hold themselves back.

And people related to each other through roles, which he calls grounding authority: a request was made to a role and from a role, so authority stopped being a property of the person who used to be in charge. Without that, he found, authority goes on "following leaders like a shadow even after it has been formally redistributed".²⁴

He's equally exact about why it breaks down. Keeping roles up to date lost out to urgent work. The complexity pushed people back to asking whoever was senior: one person found forty people to bring along where there had been one or two. Talking through roles felt unnatural and cold to some, and others simply wanted a boss.

The same former managers kept the leading roles, so role and person blurred. Anything left unwritten fell back to rank. And the role that caught the unassigned work never got a clear edge.

Lee's judgement is sober. Without constant effort it's like swimming against a current, and he closes by saying decentralisation "may never be fully achievable".²⁵

What we chose, and why. Branches are our version of this, and we reached it for the same reason: when nobody has written down which part answers for what, somebody ends up handing out the work. Each branch has a purpose, the things it answers for, and its working practices, written down. Its purpose and what it answers for are changed by proposal in the forum around the branch's parent, by consent, with a concern being a reason the change would do harm and not a preference.

And people talk branch to branch. To raise something with another part of the organisation, a seat holder chooses which of their branches they're speaking as and which branch they're speaking to. There's one conversation for each pair of branches, anyone in the organisation can read it, and an ask made there becomes a piece of work on the other branch. There's no private chat between people.

How we think it lines up. On bounding, closely: written, public, revisable by a low-threshold process, and working both ways. On grounding, more closely than we knew when we built it. Speaking as a branch to a branch is what Lee calls relating through roles, made the way one part asks another for something.

Two of our choices bear on his frictions. A branch is shaped around a piece of work with its own purpose, and the organisation's purpose sits at the root, so the "why" is written beside every "what". And keeping the structure current is lighter here, because it's a by-product of working: the branch is where the work arrives, so there's no separate map to keep up.

Where we might be wrong. Lee's frictions are the costs of doing exactly what we've built in, so they're ours to expect. Speaking as a branch may feel as cold here as speaking as a role did there. We've added a gathering that's kept off the record, and a companion whose advice is there to be heard, never obeyed. Whether a system can be made warm by design is a fair thing to doubt.

The same people may keep the same seats for years, and a seat that never changes hands blurs into its holder. Nothing in the design moves seats on.

On the forty people, there are two parts and we've answered one. Everyday work that crosses branches doesn't follow the tree. It's routed to whichever branch answers for it, and an ask from one branch becomes a piece of work on the other. Where a customer's matter ends up split across branches, that shows as a tension.

The harder part is building something new across the organisation, which needs many branches to move together. Our answers are the shared reading, tensions that climb to a forum that can act, and the Field where people sense together. None of them has been tested against a real cross-cutting push.

In his dissertation Lee also warns that people may stop noticing written roles over time. We think that risk belongs to roles kept as a map beside the work, which people have to remember to consult. A branch isn't a map. It's where the work arrives, so there's nothing to forget to look at. If what a branch says it answers for goes stale, work stops finding it and shows up with no home, which is how the staleness gets noticed.

6. Conflict between the parts

Our understanding. The Danish rehabilitation service had the tools an advocate would list: an advice process, written role agreements, peer mediation, and a council of representatives from each team to watch performance, settle disputes between teams and share information. Inside teams it worked. People reported more shared responsibility and a safer relationship with what management remained.

Between teams it broke. A change the staff had chosen themselves made the teams depend on each other for the flow of people coming in for treatment. Two teams fell out over it. Nobody would confront a colleague in another team, and no representative wanted to go back and disappoint their own.

The council simply stopped. One manager's summary was "Each team developed their own rules of the game". Teams kept the peace by keeping people in their own programmes, at the expense of those waiting for treatment.

And none of it was said out loud. The hopeful story was told on official occasions and the rest only in private, so what people knew never reached anyone who could act.⁷

The authors' conclusion is that decentralising managerial authority "can be an impossible endeavor" where teams depend on each other and there's no working way to handle conflict between them. Their recommendation is that handling disagreement has to be built into the organisation, and that the work of spanning boundaries should be everyone's, not a job for a few representatives.²⁶

What we chose, and why. We keep three things apart that the Danish council was asked to do at once.

The forum is for structure. It decides what a branch is for, what it answers for and who is seated. It isn't where a quarrel is settled, and we don't ask the stewards of sibling branches to rule on each other.

The Field is for raising things formally. Any member can put a tension or an idea to a part of the organisation, or to all of it, as a sensing. People answer in their own words and mark what lands with them, blind, with no names or counts until it closes. What comes back is a collective sense of where things stand and which way to go.

The Clearing is for airing. It's a gathering with a form to it, kept off the record, where what's been building can be let out before it hardens into blame.

Under all three sits the organisation's agreement on how its people are here together, which each member agrees to on joining. The starting text is the fifteen commitments of conscious leadership. It's meant to be used, as a filter for what gets raised and a compass for how.

And what passes between branches is visible to everyone. The conversation between any two is open, an ask becomes a piece of work on the other branch, and a slipped promise rises by itself. The organisation's reading shows where work keeps bouncing between two branches or is repeatedly sent where it isn't held, as a tension for people to look at. A tension nobody answers climbs to the next forum every fourteen days.

How we think it lines up. On visibility, better than the Danish case. There, a colleague could quietly block the flow for a month and the only remedy was to knock on her door. Here the work that's waiting, and on whom, is in front of everyone, and nobody has to accuse anybody for it to be seen.

On the authors' main recommendation, that handling disagreement be built into the organisation, we think the design has the parts: a formal place to raise a tension, a separate place to air it, and an agreement on conduct to hold both. And the Field answers their point about official stories. A sensing is blind, so what people will only say in private has a way to be said.

Where we might be wrong. The Danish service had a peer mediation procedure too, and nobody used it. Having a place for conflict isn't the same as people going there. The staff could all see the problem and still said nothing, and a blind sensing still needs someone to call it.

There's also a gap between sensing and settling. The Field can show that two branches are at odds, and what the organisation makes of it. If the answer is a change of structure, that goes to a forum. If it isn't, nothing in the design says who brings the two together or what happens when they don't agree. The authors ask for boundary-spanning to be a shared obligation. Ours rests on whoever feels the tension choosing to raise it. All of this is untested with real people in a real quarrel.

7. Who takes up authority, and who is harmed

Our understanding. Lee's experiment is the hardest evidence on this page. In the State of Washington's technology agency, groups covering about half of five hundred people opted in; some were given a group-level version of Holacracy for twelve months and the rest carried on as before, with a training for managers on encouraging participation.

On average it made no difference to how empowered, engaged or satisfied people were. But it wasn't nothing for everyone. The effect was more positive for people their managers had rated as high performers, for people who'd wanted it, and for people in groups where it already felt safe to speak up, and it turned negative for some of the opposite groups. The pattern was clearest for performance: high performers gained on all three measures and low performers lost on all three.

Lee's conclusion is that decentralising authority "can do harm as well as good", and that it's "not a panacea nor for everyone".²⁷

We should say what the experiment wasn't. Groups chose to take part, and only nine of the thirty-nine were assigned at random. The groups sat inside an agency that otherwise kept its hierarchy, which Lee calls an artifact of the design; by his own definition, we'd add, they weren't a self-managing organisation. And the agency was still unsettled after a merger, with a budget shortfall found part-way through.

Valkiainen's study suggests why being given authority isn't enough. In the firm he worked with, people who plainly had authority still didn't use it. Told a decision was theirs, "they look around the room for others to validate it". He reads this as learned helplessness: years under hierarchy leave people passive even when control is on offer. And he found its mirror in the former managers, a reflex to step in, guide and decide. The two feed each other: "deference invites intervention and intervention strengthens deference".²⁸

He names three things that helped people unlearn it. Clear roles made authority visible. Repeated chances to act from a role's purpose let people find that it held. And regular forums, where tensions were voiced and requests made, gave people somewhere to carry the discomfort. Clear roles, he says, are a precondition and not enough alone. The process was slow, uneven and reversible.

The same people turn up in the other cases. In the Danish service, those who didn't want the change were sidelined. In the French firm and in Lee's company, some workers wanted a boss back.²⁹

What we chose, and why. Little in the design was chosen with these findings in front of us, and we'd rather say so. But one choice bears on them more than any other.

Lee's groups were handed a new process to run: new meetings, new rules for changing roles, and the job of working out what to do with their new authority. He gives that added work as one reason some people ended up worse off.

Here the mechanism is in the operating system, so it doesn't depend on a meeting being held or run well. A proposal can be raised, advised on and decided without one. That isn't a substitute for meeting. People can still sit down together, and there's a lot a meeting gives that a system can't: working a tension through, and coming to understand each other.

Work arrives already routed to the branch that answers for it, often with a reply drafted beside it. A decision arrives as a proposal to consent to or raise a concern about, with its evidence attached. So what reaches a person is close to: here's what's yours, and here's the next step.

Our bet is that this suits people who like clear work in front of them as well as those who like to roam, and that it asks less of the proactivity Martela says self-management needs.

Other choices bear on it too. Joining is by personal invitation from someone who holds a seat, so nobody is put into an organisation like this who didn't choose it. Seats are offered and not assigned, and a member can belong without holding one.

Beneath pay an organisation can set an equal floor that isn't earned by work; it starts at nothing, and a floor no forum can vote away is something we're building towards. And before The Grove goes live, the way in is a sandpit: time spent in a simulated organisation, so people can find out whether they want this before anything depends on it.

How we think it lines up. Lee's three conditions are interest, performance as rated by managers, and a safe climate. On interest, the design selects for it, which is what Buurtzorg did: many of its nurses left other employers to join.⁶ On performance, we're betting that work which finds its branch, and a companion beside every seat, lower what self-management asks of a person. On climate, we have an off-the-record space and no published scores, and that's all.

Against Valkiainen's three, two fit. Clear roles are the design's foundation. And work that arrives with its next step is a steady supply of small chances to act, with the sandpit as a place to practise before anything is at stake.

Where we might be wrong. His third condition is a warning about how our main bet could be misread. Lee's experiment found the new meetings a burden. But Valkiainen found the regular meetings were what made autonomy bearable: a place where people had to speak, heard each other, and learned that tensions could be said.

The operating system makes meeting unnecessary for deciding, and an organisation could take that to mean it's unnecessary altogether. Nothing in the design asks people to meet, or notices if they've stopped. The Field and the Clearing are the nearest things it offers, and neither is where decisions are made. If the support Valkiainen describes matters as much as he found, it rests on each organisation choosing to keep meeting.

One feature may feed the loop he describes. When a promise is at risk, someone holding a branch above may take the step. That keeps the promise. It's also an open invitation to the reflex to step in, and each rescue teaches the person below to wait for the next one.

And the wider doubt stands. Lee cites the view that self-management needs "a higher level of psychological development and interpersonal skill",³⁰ and his experiment suggests it amplifies whatever is already there. People in a team that feels unsafe before are likely to end up less engaged and less satisfied, and no rule in software repairs that.

Making the work simpler may not be what his low performers needed; we don't know what they needed. A companion might equally let capable people run further ahead. Selecting for people who want it is honest, but it narrows who this is for.

And one tempting path is the one Lee's experiment warns against: a self-managing team inside a company that otherwise keeps its bosses. His result says to be very careful there.

8. Crisis

Our understanding. Every organisation that decentralises will sometimes need someone to decide fast. Lee says so plainly, and adds that "hierarchy and centralization are not synonymous": authority can be gathered into a role with edges, without bringing rank back.³¹

Lee and Young-Hyman watched four decentralised organisations do it during the pandemic. All four centralised to about the same degree, in the same two areas: rules for how people worked, such as safety and working from home, and money, with spending freezes and, in three of them, pay cuts or layoffs. Three kept their people's commitment and one lost it. The difference was how.

In the three, the authors found what they call a democratic deviation: "leaders and workers collectively authorizing centralization and then leaders enacting that authority with openness and transparency".³² In practice the extra authority was either already written into the rules, and leaders pointed to the rule when they used it, or it was created by the organisation's ordinary process for changing who decides, where others could object and amend.

It was limited in scope and in time. In one case a worker on the board saw it as his job to watch the end date.

And in two of the three it was ended out loud. One organisation formally closed a measure that had already fallen out of use, because, its chair said, "Symbols and rituals are important."

The fourth ran on principles and custom, with nothing written about who held what or how that could change. Its leaders began meeting almost daily and took decisions back, quietly, to protect the culture. People found out when requests were refused. One summed up the new rule as "Experts make decisions, asterisk, except in times of financial difficulty". Then deference spread well beyond what the leaders had taken: people began asking permission for things nobody had centralised. There was no end to it, because it had never formally begun.³³

The authors' conclusion is that what has to stay decentralised isn't any particular spread of authority. It's the power to decide who decides. Rules for that have to exist before the crisis, they have to need people's consent, and they only count if they're used. And they found this could hold without shared ownership.

What we chose, and why. The first half is built, and it's the heart of the design. The operating system has exactly the two levels of rule the authors describe. Seats and what branches answer for say who holds authority day to day. And every change to those goes through a proposal in a forum, by consent, on the record. Nobody can take authority quietly, because there's no way to change who holds what except in the open.

Most of what a crisis needs is also possible without any of that: people can act on work at once, reply, hand things on, and spend from their own branch's budget on their own word.

The second half isn't built. We've sketched what we call the fire alarm, and whether to build it is still open.

Someone pulls it, naming the condition, the part of the organisation it covers, and who takes command. For a few hours that person acts alone within that scope. It can never change the purpose, move money to a person, take away anyone's right to leave, rewrite how decisions are made, or extend itself.

It expires on its own, and calling it off is easier than any ordinary decision. When it ends, the commander accounts for everything they did, and the forum takes each act as agreed or not.

How we think it lines up. The sketch matches the three cases that held, point for point: authority that exists in the rules before it's needed, named when it's used, limited in scope and time, exercised in the open, and ended explicitly with an account. And the fourth case's failure, a quiet takeover with no start and no end, is the thing the operating system makes hardest to do.

Where we might be wrong. It's a sketch. Until it's built, the honest answer to "what if the building's on fire?" is that the people in it act and the operating system catches up, and those are exactly the conditions under which the fourth case went wrong.

The alarm as designed runs for hours. One of the pandemic measures in this study was set for thirty days, and for anything that long our only route is ordinary proposals on ordinary clocks, which may be too slow, or may push people to act outside the system.

And the study found that being authorised by others is what made leaders careful. An alarm one person can pull on their own word has less of that than a proposal others have had the chance to amend.

9. The centre, moving across, and where it fits

Our understanding. In the best-known cases, two things were decentralised least. Strategy stayed with the chief executive at Zappos and Morning Star, and Valve's executives made the biggest calls. Hiring and firing weren't decentralised at Zappos, were driven by work groups after consulting influential leaders at Morning Star, and ended with the chief executive at Valve.³⁴

Buurtzorg's teams do their own hiring, but top management holds strategy, the design of the organisation and dealings with insurers and government. Its coaches are worth noting. Each supports forty-five or more teams, and it's made explicit that they have "no decision-making authority".⁶

In Lee's company, six months in, the five most senior leaders set strategy at an off-site with no role authorising it, and then wrote one. Some people took that as proof nothing had changed. Others welcomed it: "It takes the power out of the person and moves it to a role that anyone could fill".³⁵

On moving an existing organisation across, the French dairy firm is the cautionary case. Its former middle managers were asked to build the thing that removed them. The researchers call them, in a borrowed phrase, "knowing architects of their own demise". The chief executive, the founding family's eldest son, left them to find their own way.

Their resistance grew from private cynicism to a secret group to an open campaign, using the two outside audits as their evidence. They rebuilt the missing records in secret and took them to the teams and then to the chief executive, who brought managers back, in a flatter hierarchy than before. His verdict afterwards was "The mistake I made was removing the hierarchy".³⁶

The researchers' point is that the former managers' criticism was neither invited nor used. That it found its own way out is our reading.

On fit, Martela names three conditions: units that don't depend much on each other, work that's tailored and not standardised, and skilled, motivated people. Buurtzorg has all three, and it scales by adding small teams that barely need one another, splitting any that pass twelve people. The Danish case is what happens without the first.

Elsewhere the record is thin. GitHub ran without bosses to a few hundred people and then brought managers in. Zappos's staff turnover rose by half the year it adopted Holacracy in full.

The Stanford review finds that decentralisation tends to reverse when the decisions a growing organisation needs overwhelm its way of making them, and that a spread of authority on paper can hide concentration in fact. Martela's own verdict on whether this is the future is that it's "too early to say".³⁷

What we chose, and why. We decentralised both of the things the cases held closest. Purpose is held at the root and changed only by proposal in the root's forum; a founder's strategy enters as bets that the organisation tests. People join by personal invitation from someone who holds a seat, and in a newcomer's first month other seat holders can question the invitation.

Seats are given and handed on by forums. Nobody can be fired by a person: peers withdraw their trust, and a forum can pass a seat on. Anyone can leave at once, with no permission.

We did this because a design that keeps strategy and membership at the top has kept the two decisions that most make someone the boss.

How we think it lines up. Here we've gone further than most cases in this research on strategy, which means we have the least evidence behind us. Lee's Strategy Council is close to how we'd want it done: strategy as a named seat with edges that anyone could hold, decided in a forum, and never as a room people find out about afterwards. On dissent, the design gives what the dairy firm lacked. Anyone can raise a tension, anyone can speak in the Field, and a tension that's ignored climbs.

The design also lacks what the dairy firm had: one man to persuade. The former managers won by convincing the chief executive that removing the hierarchy was a mistake, and he put it back. There's no such route in the operating system. Nobody holds the power to restore a chain of command. Structure changes only by proposal in a forum, and the operating system has no way to make one person another's manager.

Where we might be wrong. Strategy may resist decentralising for good reasons. Martela notes it's hard to imagine a division voting to abolish itself. A forum weighing bets may be slower and more timid than one person with a view.

On moving across, we have almost nothing. We've rehearsed a business that already runs being laid in on its first day, but never one arriving with managers. An existing company arrives with managers, and this research says they're where it's won or lost: their urge to step in, their knowledge of what the outside world demands, and their reasonable fear of having no place. Nothing in the design says what becomes of them. It's the place where we think the design is thinnest.

On fit, everything we've tested is simulated, and most of it small, which is the easy case by Martela's conditions. Tightly linked work, low-skill work and thousands of people are all untested. Buurtzorg's lesson is that scale came from modularity, and we don't know whether branches give the same thing. And we have nothing like its coaches: experienced people with no authority who help a team that's stuck. Wise-Eye isn't that. A network could offer it, organisation to organisation, but nothing in the design provides it.

Where things stand

Where we're least sure

These are the choices we'd most like people who do this research to push on.

1. Conflict between branches. We make it visible, give it a formal place to be raised and a separate place to be aired. The Danish case had a mediation procedure and nobody used it. And nothing in the design says who settles a quarrel that isn't about structure.
2. Moving an existing organisation across. The design assumes an organisation that starts here. The research says former managers are where a transition is won or lost, and we say nothing about them.
3. Meeting left to choice. We put the mechanism in the system so deciding doesn't depend on a meeting. It isn't meant to replace one. But one study found regular meetings were where people learned to use their authority, and nothing in the design asks an organisation to hold them.
4. Peers who won't act. The work is open and peers can withdraw trust or hand a seat on, with no published score. Two cases show people who couldn't bring themselves to sanction a colleague. If that's the norm, our sanctions exist mostly on paper.
5. A sealed core. What each branch answers for is the organisation's own and guesses are marked, but an organisation can't read the code. The research on platforms says that's where hidden hierarchy lives.
6. Strategy and membership held by forums. The best-known cases kept strategy at the centre. We may have decentralised the thing that most needs a single mind, or the thing that most makes a boss.
7. All of it on simulated organisations. Everything here is design and rehearsal. This research is a handful of cases and one experiment, and it's still far more evidence than we have.

And if we've misread this research anywhere on this page, please tell us. We'd rather be corrected than be generous to ourselves.

If you'd like to go further, these chapters say more: Chapter 7, The spark, Chapter 9, The branch, Chapter 11, Making decisions, Chapter 12, Trust, Chapter 13, Work that finds its way, Chapter 15, Promises, Chapter 17, Money and value, Chapter 19, The read, Chapter 20, Holding the whole without a centre and Chapter 22, The co-pilot.

Notes and further reading