Why Adopting AI Requires Adopting Complexity
A few months ago, I wrote that AI adoption is really about complexity adoption. Since then, the conversation has shifted, and ongoing developments are making that argument harder to ignore.
The question is no longer whether organisations should adopt AI. In many cases, that decision has already been made. What has become more urgent is a different question: why isn’t it working? Despite the challenges being identified and framed around skill gaps, data quality, and governance policies, pilots still stall. Governance gaps persist. Tools are introduced without the structures to support them. And with the technology improving at a rapid pace, the organisational question gets harder to set aside.
To understand why, the distinction between automation and AI matters more than most give it credit for.
Automation is an instruction that delivers a predictable outcome for a process that can be defined in advance. You know the steps, the exceptions, the desired outcome. The logic can be translated into code and operationalised through rules, systems, and controls. Automation is designed for repeatability within linear and fixed constraints, and its value comes from being faster, cheaper, and more reliable. AI augmentation and agentic approaches create value under different conditions. They become useful where rules are incomplete, where inputs are ambiguous, where judgment matters, and where iterative work with probabilities, patterns, and examples gets you further than any fixed instruction can. Automation thrives on stability; augmentation and agents become most valuable when reality is more complex than the process map, a distinction that maps directly onto what cybernetician W. Ross Ashby called requisite variety: a system can only regulate complexity up to the level of its own internal variety.
Most organisations, however, are still designed as systems that process instructions. Structures, reporting lines, management logic, and performance models are purposefully built to reduce variation, increase control, and optimise predictability. That is not a flaw but a rational response to scale, efficiency pressure, and operational risk. And it explains why the majority of organisations and their AI initiatives currently focus on process automation. From a business perspective, that focus makes sense. If margins are under pressure, using AI to reduce cost and improve efficiency is a valid move. But there is a strategic limit to that logic. Optimising current operations for efficiency improves the margins of what created value yesterday. It does little to build the organisational capability needed for what customers, markets, and competitive conditions will require tomorrow.
Within that gap resides the real tension. Rational near-term AI deployment doesn’t necessarily build lasting AI capability. Adoption and transformation are not the same thing. Adoption means the tools are being used. Transformation means an organisation has fundamentally changed how it operates, and is able to achieve better outcomes through these tools. Unsurprisingly, AI does not simply slot into how most organisations work today. In many cases, they are actively structured against its effective use. Not because leaders don’t care, and not because the technology is immature, but because the organisation is designed as a system that rewards control, short-term delivery, and complexity reduction, and penalises the friction that profound change creates.
The above outlined dynamic, the structural bias towards optimisation, is also why the problem runs deeper than most current AI narratives suggest. Being a dynamic, it is not a single problem, or primarily a “knowledge problem”, but inherently a systems design problem. Disrupting your own business model from inside a scaled organisation — with years of infrastructure, processes, commitments, and incentives built around it — is expensive, slow, and often career-limiting in ways that optimising the current model simply is not. So even where the need for change is visible, the organisation often remains more capable of defending its existing logic than of adapting it, a dynamic Clayton Christensen documented in a different context: rational managers in well-run organisations making decisions that collectively produce the wrong strategic outcome.
To frame the systems design problem in a more tangible way, consider how goals, metrics, and reporting actually work in the majority of today’s organisations. Objectives cascade top-down, even when OKRs or similar frameworks are in place. On paper, this looks like alignment. In practice, it often functions as a one-way system of confirmation. Teams close to operations usually know where processes create friction, where customer needs exceed existing offers, and where new possibilities are emerging — but the reporting system doesn’t ask for that. It asks for data that confirms the current plan is on track. So that’s what gets presented. Information travels upward in forms the system can absorb — which usually means filtered, simplified, and made safe. While the knowledge of how to improve exists inside the organisation, the system isn’t built for wanting it.
AI beyond automation creates value at the intersection of ambiguity and operational reality. To be effective, it depends on organisations being able to surface weak signals, work with incomplete knowledge, and translate local insight into changed decisions, structures, and new offerings. If the mechanisms for that don’t exist, AI gets pushed back into deterministic use cases that fit the existing model. So a general lack of this adaptive capacity — the condition under which AI becomes meaningfully useful — is why better tool selection, more pilots, or another AI task force rarely solve the core problem on their own. They may generate momentum and produce isolated wins, but they don’t build the organisational capacity to absorb and work with complexity. Most programmes and their respective metrics are focussing on achieving the first while assuming it will produce the second.
Increasingly, the required adaptive capacity brings the challenge back to organisational design. Structural answers to the following questions might indicate whether AI adoption is facing the foundational organisational conditions that determine whether it can succeed.
Can operational insight reshape business strategy, priorities and decisions, or does it stay local and only travel upward in forms the organisation can absorb?
When AI surfaces something ambiguous or unexpected, is there a clear decision logic in place, and who has the mandate to act on it?
Do your governance structures enable responsible judgment under uncertainty, or do they only function when the answer is already known?
The real challenge for organisations isn’t adopting AI. It’s being able to operate within complexity. What AI augmentation and agentic workflows require is not only new tooling, policy, or skills. They require organisations to become better at working with complexity itself: embedding continuous feedback and learning loops, operating models that can adapt, governance that holds under ambiguity and competing pressures, and collaboration systems that let insight reach the people who need it. Once that organisational foundation exists, many things that once seemed too complex to attempt become tractable. Not because AI got smarter, but because the organisation did.
Within this context, the key question to ask isn’t “how do we adopt AI, fast?” but the far more uncomfortable one for leadership to address: “how ready are we to work with complexity and distributed judgment?”
And beneath that sits the even harder one: who in your organisation actually has the mandate to make that possible?
References
W. Ross Ashby — Introduction to Cybernetics (1956). The Law of Requisite Variety: a system must match the complexity of its environment to regulate it effectively. The theoretical basis for the argument that organisations designed for instruction-processing cannot govern complexity.
David Snowden & Mary Boone — A Leader’s Framework for Decision Making (Harvard Business Review, 2007). The Cynefin framework distinguishes between complicated domains where best practice applies and complex domains where emergent practice is required — the conceptual basis for the automation vs. AI distinction.
Chris Argyris — Organizational Learning (1978). Single-loop vs. double-loop learning, and organisational defensive routines — why teams present information that confirms the current model rather than challenges it.
Clayton Christensen — The Innovator’s Dilemma (1997). Why rational decision-making inside successful organisations systematically produces the wrong strategic outcome — applied here to the incentive structures that work against meaningful AI adoption.
Andrew Ng — writings and lectures on AI transformation (deeplearning.ai, ongoing). The most consistent practitioner voice on why AI adoption fails at the organisational rather than the technical level — and what closing that gap actually requires in practice.
European Union — EU AI Act (2024). The regulatory framework governing AI deployment in high-risk environments across EU member states. Establishes governance, transparency, and accountability requirements that make the organisational design questions raised here not only strategic but legally material.
MIT Sloan Management Review — AI & Organisational Design (2023–2024). A body of current practitioner research on the gap between AI capability and organisational readiness — bridging the theoretical foundations above and the operational realities organisations face today.