Most leadership frameworks were designed for a world that behaves like a clock. You set the gears in motion; the output follows. Hire consultants, implement best practices, run the plan. But organisations don’t work like clocks. Neither do markets, cities, supply chains, or societies. They work like ecosystems — and that distinction has immediate, practical consequences for every decision you make.
The field that explains the difference is complexity science. Specifically, the theory of complex adaptive systems (CAS). It has been developed over four decades at places like the Santa Fe Institute and refined by management thinkers like Dave Snowden and Ralph Stacey. Its relevance is sharpest now: AI is adding a new class of non-human agents to every system we manage.
Here’s the core idea, then we’ll get practical.
What Makes a System Complex?
A complex adaptive system has three defining features.
Many interacting agents. The system is made up of actors — people, firms, algorithms, cells — each acting on local information, pursuing their own goals. No single actor controls or even sees the whole. The economy is billions of consumers and millions of firms. A hospital is doctors, nurses, patients, administrators, and suppliers, all continuously adapting to each other.
Emergence. The system’s behaviour is not the sum of its parts. Traffic jams emerge from individual driving decisions, but no single driver chose to create one. Financial panics emerge from individually rational decisions by banks. Organisational culture emerges from thousands of daily choices by people who were never trying to create culture. Emergence means the system has a life of its own that you cannot predict by studying its components in isolation.
Non-linearity and feedback. Small inputs can produce massive outputs; large inputs sometimes produce nothing. Outputs feed back to change the agents, who then change the system again. Toyota introduced a single rule — any worker can stop the production line — and the culture of quality that emerged from millions of small interventions over decades became one of the most studied management systems in history. The rule was simple. The feedback loops were powerful and long-running.
The physicist Murray Gell-Mann, a Nobel laureate and Santa Fe Institute co-founder, put it plainly: complex adaptive systems are systems that learn. They are not passive. They respond to what you do to them.
The Framework Leaders Actually Use
Dave Snowden’s Cynefin framework is the most practically useful tool to emerge from CAS theory. It maps problems into five domains.
Clear: cause and effect are obvious. Best practices apply. Follow the procedure, train the staff, measure the output.
Complicated: cause and effect require expertise to unpick, but they are knowable. Engineering, finance modelling, medical diagnosis. Hire experts, analyse the situation, act on their recommendation.
Complex: cause and effect can only be understood in retrospect. No expert can reliably predict the outcome before you act. Snowden’s prescription for this domain is probe, sense, respond — run small, safe-to-fail experiments, watch what actually happens, amplify what works and dampen what doesn’t.
Chaotic: no discernible cause-and-effect. Act fast to establish a foothold of order, then move toward complex or complicated.
Confused: the centre, where you don’t know which domain you’re in. The most dangerous place, because you default to the approach you’re most comfortable with — which is usually wrong.
The critical insight: most leadership training treats everything as complicated. Hire the consultant, build the model, execute the strategy. That works when your problem is genuinely complicated. But if your problem is complex — and most strategic, organisational, and societal challenges are — applying complicated-domain thinking is not just ineffective. It gives you false confidence in a map of terrain that keeps moving.
The 2008 financial crisis is the canonical case. Every major bank had sophisticated risk models built on complicated-domain assumptions. The models mapped each bank’s own portfolio risk. None of them modelled the system-level feedback loops — that when enough banks tried to de-risk simultaneously, they would create the contagion they were each individually trying to avoid. Rigorous complicated tools, applied to a complex system, with catastrophic results.
Where Most of Your Hard Problems Actually Live
Ralph Stacey approached the same problem from a different direction. He plotted organisational decisions on two axes: how much certainty exists about what will work, and how much agreement exists about which goals to pursue.
Most operational decisions live in a comfortable zone — high certainty, high agreement. The right answer is largely knowable, and everyone wants roughly the same outcome. Standard management works fine there.
Most strategic decisions live in the middle — what Stacey calls the zone of complexity. Certainty is low because the environment is shifting, the technology is new, competitors are adapting. Agreement is also low because different stakeholders have genuinely different interests and mental models. In this zone, management-by-prescription fails. You cannot plan your way out of complexity with a more detailed plan.
One important caveat: Stacey himself later distanced from the matrix he inspired, arguing it oversimplified reality by treating certainty and agreement as stable axes when they are themselves in flux. He was right. But the underlying message holds: in conditions of low certainty and low agreement, your job is not to impose the answer. It is to shape conditions from which good answers can emerge.
What This Means in Practice
These are not abstract principles.
Stop optimising for pure efficiency. Toyota’s 2011 lesson: after the Tohoku earthquake, its just-in-time supply chain — optimised over decades to eliminate waste — seized up almost immediately because the system had been stripped of redundancy. Recovery happened through local adaptation by individual plants and suppliers, not through central coordination. Complex systems need slack. Efficiency and resilience are in tension, and you need to decide consciously which you are buying at the expense of the other.
Distribute decisions to where the information lives. In a CAS, local agents have information the centre does not, and they have it faster. Centralising decisions makes you slower and, perversely, less informed. The role of leadership in a complex environment is not to be the decision-maker. It is to set the context — norms, incentives, information flows, boundaries — from which good decisions can emerge closer to the ground.
Run experiments, not programmes. When you do not know what will work, a large-scale rollout is a large-scale bet. Small, fast, parallel experiments let you learn from the system rather than impose on it. This is what Snowden means by safe-to-fail: design interventions whose failure teaches you something without bringing down the organisation.
Watch the edges. Emergent change shows up at the margins before it reaches the centre. The teams running unusual experiments. The customers using your product in ways you did not design for. The competitor quietly changing its economics. These are the system’s early warning signals. If your information flows are structured to suppress them — as many hierarchies are — you will be perpetually surprised.
Why AI Compounds All of This
AI does not simplify complex adaptive systems. It adds faster, more numerous agents to them.
A model deployed in a financial market adapts to the market’s patterns — and in doing so, changes those patterns. Other models adapt to the new patterns. The feedback loops now run at machine speed. Human observers are working in retrospect before the next cycle has completed.
In organisations, AI agents are being embedded in workflows, making recommendations, filtering information, allocating resources. Each is an agent responding to local signals. Their aggregate behaviour will produce emergent effects that no one designed, on timescales faster than human review cycles can track.
This is not an argument against deploying AI. It is an argument for deploying it with CAS thinking rather than clock-world thinking. The question is not only what does this system do — it is how does this system change the system it operates in. What feedback loops does it create? What behaviours will it amplify, and what will it dampen? Who can see the emerging patterns, and how quickly?
The organisations that navigate the next decade well will not necessarily be the ones with the best AI models. They will be the ones who understand that an AI-enabled organisation is a more complex adaptive system than the one before it — more agents, faster feedback, higher-stakes emergent effects. The frameworks to think about this clearly already exist. Whether you apply them before your next major decision — or after it has already produced effects you did not anticipate — is the choice in front of you.
Sources: John H. Holland, “Hidden Order: How Adaptation Builds Complexity” (1995); Dave Snowden and Mary Boone, “A Leader’s Framework for Decision Making” (Harvard Business Review, 2007); Cynefin framework documentation, The Cynefin Co; Santa Fe Institute CAS research; Toyota Production System / CAS case study (Galaxiez, 2024); Stacey Matrix overview, NCCMT/Praxis Framework.