The Law of Self-Similarity

Written by

Sergio Caredda

Posted on

May 28, 2026

1. Structure replicates

The organisation chart you draw for your top team is the template the next layer will reproduce. Not by instruction, but by pattern-matching: the people who built the layer below were trained by, promoted by, and observed the layer above. When they faced structural choices of their own, they reproduced what they had learned.

This is the insight behind Conway's Law, originally framed for software: organisations produce systems that mirror their communication structures. Researchers call this the mirroring hypothesis — technical dependencies mirror organisational ties because mirroring conserves cognitive resources. The law is a specific application of self-similarity — your systems are fractal expressions of how your organisation actually works, whatever your design documents claim.

The implication is uncomfortable. If your top team makes decisions through informal bilateral channels rather than formal governance — even while running workshops on collaborative leadership — then every layer below will learn that informal bilaterals are how things actually get done. The formal channels are decoration; the fractal is built on what is actually practised.

2. Behaviour replicates — and the generating rule

This is the concept this law adds that the two preceding laws in this series do not fully capture.

A fractal is not complicated. One recursive instruction — the generating rule — applied iteratively produces a pattern of arbitrary complexity. Think of a tree: the same branching rule applied at each fork, from trunk to the finest twig, produces the whole structure. The same applies to organisations.

The generating rule is not your values statement. It is not your leadership competency framework. It is the set of behavioural instructions — formal and informal — that govern how decisions are actually made, how conflict is actually resolved, how people are actually rewarded when the quarterly numbers are under pressure.

If the generating rule at the top is "protect your function and win the internal competition," then function-protecting and internal competition will replicate downward. You can publish as many values about collaboration as you like. The fractal is generated by what leaders do, not by what they say.

This is why behaviour change programmes that focus on individuals rather than on the generating rule fail so reliably. You are reshaping individual outputs while the rule generating them remains intact. As soon as pressure arrives — a budget cycle, a reorganisation, a missed target — the generating rule reasserts itself. The individuals snap back. The fractal reconstructs.

3. Culture replicates

In every culture change engagement I have led — at Campari, and at VF Corporation before — the most reliable predictor of failure was not poor communication, insufficient resources, or absent executive sponsorship. It was the generating rule at the leadership level that remained unchanged while everything else around it was redesigned.

We would run workshops, produce new behavioural frameworks, align HR processes, train thousands of people. And then watch those investments cancel out at the next leadership team meeting, where the same decision-making patterns played out as before. New vocabulary, same fractal.

Culture is fractal. The pattern at the top produces the pattern everywhere below. The only way to change the culture is to change what is generating it — and what generates it starts at the senior level, specifically in the informal behaviours that are visibly rewarded, tolerated, or punished in moments of genuine pressure. The generating rule is not the same thing as culture: culture is the observable output; the generating rule is the mechanism producing it. Changing the output without changing the mechanism is the definition of a failed culture programme.

This applies at any scale: a thirty-person company in its third year already has a generating rule. It is just newer and, for that reason, more malleable.

4. What this means for design

This is where the Law of Self-Similarity connects — and sharpens — the two laws immediately preceding it in this series.

The Law of Requisite Variety established the distinction between complication and variety. Adding more structural units does not necessarily add more capability; complication and variety are different categories. Self-similarity explains why adding structure so often fails: if the generating rule is unchanged, new structural units simply reproduce the existing pattern. You have added complication; the generating rule continues to constrain variety.

The Law of Alignment [link to be confirmed] established that conformity and coherence are not the same thing — the coherence gap is real and consequential. Self-similarity explains the mechanism: conformity is what you get when the generating rule is hierarchical imposition; coherence is what you get when the generating rule is shared logic. You cannot achieve coherence by mandating conformity, because they are produced by different generating rules.

The design implication: when you want to change an organisation's output, identify and redesign the generating rule, not the output elements. Donella Meadows' hierarchy of leverage points for system change ranks changing paradigms and goals — the generating rule — as the highest-leverage intervention in any system, far above adjusting parameters and structures. Ask what instruction, applied recursively at every level, would produce the output you want. Then design your operating model to enact that instruction, and — critically — ensure that the senior leadership's visible behaviour is already governed by it. The fractal starts there.

This is harder than moving boxes on an org chart. It requires diagnosing the actual generating rule — which is never written in any document — and then changing leadership behaviour first, because nothing below it will change until above it has.

The intellectual grounding for this discipline comes from Stafford Beer's Viable System Model — developed in Brain of the Firm (1972) — which established that every viable level of an organisation must instantiate the same five management functions recursively. Beer called this cybernetic isomorphism: the same structural description applies at every scale. Patrick Hoverstadt's The Fractal Organization (Wiley, 2008) operationalised this for practitioners, treating the generating rule rather than the org chart as the primary unit of design.

One scope boundary worth naming: this law is most powerful as a diagnostic tool when a dominant generating rule is already in operation — typically in organisations that have been running for at least a few years. Very early-stage organisations, or those in the first throes of deliberate reinvention, may have competing generating rules that have not yet consolidated. In those conditions — which the Cynefin framework would classify as genuinely complex rather than complicated — the design question is which rule wins, not how to change the one that is already dominant.

Self-similarity and AI

There is a contemporary relevance to this law that is easy to miss.

As AI systems become embedded in organisational workflows, Conway's Law extends recursively. The AI-augmented processes your organisation deploys will mirror your actual decision-making patterns — not your declared ones. The generating rule gets crystallised into code.

Amazon's hiring algorithm — trained on a decade of historical hiring decisions — reproduced and amplified the gender bias already embedded in those decisions, and was discontinued in 2018 when the pattern became apparent. This is not a story about AI going wrong. It is a story about AI doing exactly what the fractal principle predicts: faithfully encoding the generating rule that was already operating.

Organisations with clear, designed generating rules will build AI tools that amplify them. Organisations whose generating rules were inherited and unexamined will build AI tools that crystallise and accelerate those patterns at scale. Which makes identifying and redesigning the generating rule more urgent, not less.

Applying the law: three diagnostic questions

Before you can change the generating rule, you need to know what it is. Most organisations do not. Three questions help:

1. What behaviour is actually rewarded when it conflicts with declared values? Not what the values say, but what happens to the person who delivers results by violating them. That outcome reveals the generating rule more precisely than any culture survey. One practical proxy: count how many decisions escalate beyond their natural level in a given week. That ratio tells you more about your generating rule than any engagement score.

2. Where does the fractal break down? A team or function that consistently produces different patterns to the rest of the organisation is either operating under a different generating rule, or has found a way to shield itself from the dominant one. Either is diagnostic.

3. What would have to be true of the top team's behaviour for the rest of the organisation to look the way you want it to? Design from that requirement, not from the org chart down.

The “Laws of Organisation Design” Collection. This article belongs to a series focused on the different principles and laws that govern organisational behaviours, decision-making and group work, and that should be taken into account in organisation design projects.

The Laws of Organisation Design

Coming Soon

Law of Synergy
Law of Minimizing Interface Complexity

Conclusion

The Law of Self-Similarity does not say organisations are deterministically trapped by their generating rules. It says the fractal is actively produced — at every level, every day — by the instruction that is actually operating. That instruction was either designed, or inherited. In most organisations, it was inherited — assembled from the behaviours of whoever happened to hold power at the moment the organisation formed its habits, and then recursively reproduced ever since.

The question the law asks is not whether your organisation is self-similar. It is. The question is whether the pattern being replicated is the one you want.