The Bossless Workplace: What the Evidence Supports, What It Doesn’t, and Why “Flattening” Is a Different Animal

Last verified: 1 October 2026 · About 14 minutes to read

Two well-known companies removed managers. Only one of them gave anyone more authority.

In September 2024, Amazon told every senior leadership team to raise the ratio of individual contributors to managers by at least 15% by the end of Q1 2025, and in the same period ordered corporate staff back to the office five days a week. Between 2013 and 2020, Zappos went the other way: it dissolved job titles and pushed decision rights down into circles and roles, then quietly walked much of it back.

Both get called “bossless.” They are opposite experiments. Most of the confusion in this topic, and most of the bad advice, comes from treating them as one.

Quick answer: does the bossless workplace work? Under narrow conditions, yes. The best-documented case, the Dutch nursing organization Buurtzorg, uses self-managing teams of up to 12 people doing work that divides cleanly by neighborhood, backed by a lot of shared infrastructure. Converting a large existing company wholesale has a weaker record: Zappos and Medium both retreated from Holacracy. And the 2025–26 “flattening” wave is not self-management at all. It removes managers while keeping authority at the top.

How I checked this. I traced each headline figure back to the most primary source I could open: company statements, the contemporaneous reporting of the underlying events, peer-reviewed reviews, and the Commonwealth Fund’s analysis of the KPMG data on Buurtzorg. Where I could only read an abstract or a secondary summary, the source table at the end says so. Evidence grades used throughout: A = primary document or peer-reviewed; B = reputable reporting of primary facts; C = secondary, company-reported or single-source; Calc = my arithmetic from stated inputs.

1. Three different things hide under one word

The non-obvious point: evidence for one column does not transfer to the others. Buurtzorg’s results say something about moving authority down. They say nothing about whether Amazon’s larger spans of control will work, and nothing about whether software can stand in for a manager. When a vendor or consultant cites Buurtzorg to justify cutting a management layer, they are borrowing credibility from a different mechanism.

2. The Zappos ledger: what the numbers actually say

The version of this story that circulates is wrong in a way that matters. Zappos did not “offer 14% of staff a buyout.” It offered every employee severance of at least three months if they decided by 30 April 2015 that Holacracy wasn’t for them, and about 14% accepted.

What this does and doesn’t prove. The 14% is partly a selection effect, since the offer was designed to make leaving easy. The second wave is confounded by an unrelated technology migration. And the 2020 retreat came after a long run, not an immediate collapse. My reading: Zappos is solid evidence that an ultimatum-style conversion of a 1,500-person company costs roughly one employee in seven within weeks, and that the full system didn’t survive as designed. It is not evidence that self-management can’t work.

Medium adds a second data point. It left Holacracy in March 2016. Co-founder Ev Williams wrote that general-purpose frameworks carry more overhead than purpose-built ones, and that Medium teams kept running Holacracy-style tactical meetings because those worked well. That detail is the most transferable lesson in this whole story: you can keep the cheap part (a disciplined meeting format) without adopting the expensive part (a company-wide governance constitution). Secondary accounts of Medium’s reasoning also point to difficulty coordinating initiatives that cross functions (grade C).

Insight: separability, not ideology. The successes I could verify involve work that splits into self-contained units: a team of nurses owning a neighborhood’s patients. The retreats involve work that cuts across units: a product that needs design, engineering and operations to move together. This is my inference from three cases, not an established finding. But it gives you a better first question than “which framework?” Ask: how much of a team’s work can finish inside the team?

3. Buurtzorg: the best evidence, and it’s smaller than the legend

Buurtzorg was founded in 2006 in the Netherlands by nurse Jos de Blok and colleagues who were frustrated with task-based home care. Teams of up to 12 nurses are responsible for the care of roughly 50 to 60 patients in a defined neighborhood. When a team grows past 12, it splits (Harvard Business School case, published on Buurtzorg’s own site).

The most careful independent numbers come from a 2015 KPMG analysis, summarized in a Commonwealth Fund case study. I could read the Commonwealth Fund’s account but not the full KPMG report.

What the table says. Buurtzorg delivers markedly fewer hours of care per client and is a low-cost home-care provider, but once nursing-home, physician, and hospital costs are added, the Commonwealth Fund summary puts its total cost per client at about the Dutch average. The famous “40%” figure is also muddier than it looks. Some popular accounts say Buurtzorg needs 40% fewer hours. The Commonwealth Fund and the Centre for Public Impact describe an Ernst & Young finding that Buurtzorg patients use about 40% of the hours they are authorized for, versus about 70% in the sector. Those are different claims, and the second is what the sources support. Turnover figures also differ by source (10% versus 15% in one, “less than half the average” in another), so I haven’t used them.

Why it works, per the documentation. It is “no management.” Look at what Buurtzorg built instead: a hard team-size cap with a split rule, a minimum productivity threshold (at least 62% of team hours spent on patient care), a shared platform that feeds teams performance data, and regional coaches. One secondary source puts coaches at about one per 40 to 50 teams (grade C). The management work moved out of people and into infrastructure.

Verdict: Buurtzorg shows self-management can run a large organization well, and the strongest result is better care with fewer hours, not a dramatic drop in total system cost. It was also built that way from day one, so it cannot tell you what happens when you convert an existing company.

4. What the research literature supports, ranked

Two corrections I had to make to commonly repeated claims:

• Project Aristotle does not support a 12-person cap. Secondary summaries of Google’s re: Work findings list team size among the factors that showed no significant link to effectiveness. Buurtzorg’s cap of 12 is Buurtzorg’s operating rule, not a finding from Google’s research.
• Gallup cuts against the bossless pitch. Its headline claim is that the manager is the dominant factor in team engagement. If true, a bossless design must replace what good managers do, not just remove the layer. The one Gallup item that does travel well is the “know what is expected of me” item, which is a clarity problem a role document can address.

Where the evidence is thinnest. I found no peer-reviewed, controlled comparison of company-wide self-management against a matched conventional firm. Anyone quoting a precise productivity gain for “mature self-managed teams” is quoting something I could not trace to a primary source.

5. The audit nobody does: what do managers actually do?

Every responsible design starts by listing the work a manager performs. This is my own working framework, built from the cases above. For each function, the question is what concretely takes its place and how good the evidence is.

The rule that bites: if you cannot name the concrete replacement for a function, that function stays with a named human. “The team will handle it” is not a replacement. Of the six, pay and conflict are where I found the least evidence, so they are where I would be most cautious.

6. AI: what’s claimed, what’s measured

The headline forecast comes from Gartner’s October 2024 predictions: through 2026, 20% of organizations will use AI to flatten their structure, eliminating more than half of current middle management positions. Three things about it:

1. It is a forecast. The window closes this year, and I did not find a published Gartner audit of whether it came true.
2. It describes flattening, not self-management. Gartner frames the benefit as lower labor cost and wider spans of control for the managers who remain.
3. Gartner lists its own risks: worker job-security fears, remaining managers overwhelmed by additional direct reports, and broken mentoring and learning pathways for junior staff.

Amazon is the clearest real-world example of that flattening logic. According to reporting from Fast Company and Entrepreneur, the company reached its manager-ratio goal by the end of March 2025 largely by merging teams and moving some managers into individual-contributor roles rather than by widespread layoffs. In late 2025, it announced about 14,000 corporate job cuts. None of this is self-management: authority stayed with senior leaders, and the company simultaneously tightened in-office rules.

The question this raises, and my answer. Can AI make coordination cheap enough to extend self-management beyond work that splits cleanly into units? My judgment, which is analysis rather than evidence: not soon. AI lowers the cost of informational coordination such as notes, status, routing, and reporting. The hard cross-unit problem is priority conflict, where two teams want the same resource, and that needs someone with the authority to decide. So expect AI to widen spans of control (flattening) faster than it enables self-management. What would settle it: before-and-after studies of cross-team decision time in firms that adopted AI tooling, comparing those that changed decision rights with those that didn’t. I found none published.

Tools, briefly. GlassFrog is the Holacracy-native option from HolacracyOne, and Holaspirit and Peerdom support role-and-circle governance; Loomio is a proposal-and-decision tool. I’ve left prices out on purpose, because published figures conflict across vendor and review sites. Treat any “AI culture monitoring” product with suspicion: surveillance is a management assumption, and importing it into a trust-based design defeats the point.

7. Should you do it? Five gates

What I would do in each case:

• Under about 50 people, mostly separable work: adopt written roles, a team cap, and regular tactical meetings. Skip the full constitution.
• A few hundred people, mixed work: pilot one self-contained unit with a coach and published metrics. Choose a unit whose outputs can be measured without a manager’s judgment.
• Thousands of people, legacy structure: the verified record doesn’t support a wholesale conversion. Flatten selectively if you must, but call it flattening and fund the middle-manager functions you keep.

I deliberately give no implementation cost or timeline benchmarks. The commonly quoted ones (dollar ranges for coaching, “20–30% productivity dips”) do not trace to a source I could verify. Measure your own: decision time on mid-stakes decisions, attrition, and how often people route around the new system and ask their old manager.

8. Frequently asked questions

What is a bossless workplace? An organization where decision authority sits with teams and roles under written rules, instead of with a management hierarchy. It is different from a flat organization, which simply has fewer layers.

Is flat the same as bossless? No. Flat removes layers and usually widens each manager’s span. Bossless (self-management) moves authority down. Amazon’s manager-ratio push is the first; Buurtzorg is the second.

Did Zappos fail? The full Holacracy system did not survive as designed. About 14% of staff left in the 2015 opt-out round and about 260 people over the following months; by 2020, Quartz reported a quiet retreat toward internal small businesses with profit-and-loss statements. The company itself continued.

Is Holacracy still used? Yes, in organizations that chose it, and software such as GlassFrog still supports it. Two high-profile adopters, Zappos and Medium, moved away from the full system.

Does Buurtzorg prove self-management works? It proves it can work in a specific model: small neighborhood-based teams with heavy infrastructure. Per KPMG’s analysis as summarized by the Commonwealth Fund, it delivers about 36% fewer home-care hours per client but only about 3% lower total cost per client.

How big should a self-managed team be? Buurtzorg caps teams at 12 and splits beyond that. That is an operating rule from one organization, not a research-established threshold.

Will AI replace middle managers? Gartner forecasts that 20% of organizations would use AI to eliminate over half of middle-management roles through 2026. It is a forecast, and I found no published verification of it.

9. What I could not verify

• Full texts. I read the HBR Beyond the Holacracy Hype abstract and author line but not the paywalled article. One of its authors, John Bunch, led Zappos’s holacracy rollout, so it is an insider’s analysis, not a neutral one.
• KPMG and Ernst & Young reports. I relied on the Commonwealth Fund and other summaries, not the original reports.
• Project Aristotle details. The list of non-significant factors comes from secondary summaries of Google’s findings.
• Morningstar outcomes. I verified the CLOU mechanism from Hamel’s article abstract, not the company’s performance.
• Not audited: I did not benchmark against the current top search results for this topic.

Sources

A note on figures: percentages marked “Calc” are my arithmetic from the cited inputs. Source figures can be revised, so check the originals before reusing a number elsewhere.