When organisations face rising costs, increasing workloads or pressure to do more with less, the instinct is to standardise, streamline, and automate. Remove duplication. Reduce variation. Make things efficient and optimised. This is the logic of productivity.
In my previous blog, I laid out the historical context for why our understanding of productivity is still deeply influenced by the logic of industrial production. The factory gave us a reductionist but intuitive model for centuries. There is an input, a process and an output. All measurable, observable and trackable. If we can reduce the time, resources or people required to produce the same output, productivity has increased. In this blog I want to talk about its consequences.
Reducing complexity, making things simple, has always been our defence against complex, unpredictable phenomena. Order and predictability make us good at managing our resources. But the very order that makes systems efficient can also make them fragile.
The more we optimise a system for the conditions we expect, the less room we may leave for the conditions we don't. This is not a new concern. For decades, thinkers across ecology, engineering, organisational theory and biology have been showing us that efficiency and optimisation can come at the expense of resilience and robustness.
The problem with making systems perfect
For much of modern science and engineering, the ideal system has been one whose behaviour we can predict and control. The problem is that this logic becomes unreliable as we move from machines to complex systems. In an ecosystem, an economy or an organisation, the relationships between components change as the system changes. The system is not simply responding to its environment; it is continually producing a new environment for itself. The distinction C. S. Holling made in 1973 between stability and resilience was important precisely because it challenged the assumption that keeping a system close to an ideal state was necessarily the same as keeping it healthy. A resilient system may need to change considerably in order to preserve what matters.
This creates a paradox at the heart of optimisation. The better we become at controlling a system under known conditions, the more exposed we may become to conditions we have not modelled. Optimisation necessarily depends on assumptions about what matters, what can be predicted, which variables can be ignored and what constitutes a desirable outcome.
Complexity theory makes a related point in a different language. Computationally, we routinely distinguish between problems that can be efficiently solved and problems where finding an optimal solution is computationally difficult. In complex real-world systems, the problem is often more fundamental: we don't even know whether the model contains the variables that matter. The danger is therefore not simply optimising badly. It is optimising with confidence inside an incomplete model of the system.
From control to adaptation
For a long time, the answer was “we should know more”. If complexity made systems difficult to predict, then the solution was to gather more information, build better models and understand more of the relationships between their components. The assumption was that uncertainty was largely a problem of insufficient knowledge. If we could observe enough, measure enough and understand enough, we could make the system predictable. This fitted remarkably well with the technologies of the time. Sensors, databases, models and increasingly powerful computers gave us an expanding capacity to measure and control.
The next response was more pragmatic: if we cannot prevent failure, we can learn from it. This is part of the intellectual territory occupied by resilience and, more radically, Taleb's antifragility. Instead of assuming that we can predict every failure, we design systems that can absorb shocks, recover and learn from them. Failure becomes information. A breakdown exposes a weakness in the system; the system is repaired or redesigned; the next failure is hopefully less damaging.
This represented a significant shift from prediction to recovery. It also fits the technological capabilities that emerged alongside it: distributed systems, fault tolerance, backup infrastructure, iterative software development and continuous monitoring all made it increasingly possible to detect failure, recover quickly and incorporate what we learned. We moved from trying to build systems that would never fail to building systems that could fail without collapsing. But failure can be too expensive a teacher. In some systems, failure is not something we can simply learn from and try again. A failed bridge, a failed medical intervention or a failed public service can have consequences that cannot be reversed.
This brings us to the most recent response: making systems as adaptive as possible. Rather than waiting for the system to break and then improving it, can the system continually adjust itself as conditions change? Can it detect that the environment is shifting and change its behaviour before the disturbance becomes a failure?
Olivier Hamant's work pushes this argument into biology and makes it particularly uncomfortable for our productivity mindset. He argues that plants are not robust because they have been perfectly optimised; they are robust partly because biological systems contain stochasticity, inefficiency, incoherence and delay. The implication is bigger than biology. What we remove in the name of efficiency may sometimes be precisely what gives a system room to adapt. Redundancy, spare capacity, variation, local decision-making and apparently unnecessary human intervention can all look like waste when judged against an ideal operating state. But in a system whose future cannot be fully predicted, those characteristics are not necessarily wasteful. They are options. They are the capacity to respond when the assumptions behind the optimisation cease to hold.
That shift matters because it changes what we mean by a well-designed system. A well-designed system does not have to be one that behaves exactly as intended. It may be one that can recognise when its assumptions no longer hold and adjust accordingly. Instead of eliminating uncertainty, we build the capacity to respond to it.
That shift is itself partly enabled by technological progress. Every technological advance expands our ability to collect information, model complexity and respond to change. Yet there is a paradox: we tend to apply new technologies to the systems we already have. We digitise existing processes, automate them and now use AI to optimise them. The technology changes, but the underlying logic does not. We are already talking about AI in planning in terms of reducing application processing times, accelerating decisions and streamlining processes. Our old productivity logic is being reborn with every new technology. The question is whether we use these new capabilities to make existing systems more efficient, or to imagine systems that work differently.
The opportunity
There is a temptation to think that the future of AI in government is about replacing administrative work with automated administrative work. I think that undersells what AI, and digital technology more broadly, could do. The more interesting possibility is to use these technologies to help organisations sense, interpret and respond to complexity, this time not just by focusing on what the technology can do, but on what humans and machines can achieve together.
That might sometimes mean automation. But it might equally mean identifying an exception, bringing the right people together, revealing a pattern, challenging an assumption or giving a frontline worker more cognitive capacity to exercise judgement. The technology becomes less about executing a predefined process and more about helping the organisation understand when that process needs to change.
This is not necessarily less productive. It is a different conception of productivity. Instead of measuring only how efficiently a system performs under normal conditions, we can ask how well it continues to function when conditions change. Instead of optimising for maximum throughput, we can optimise for the capacity to respond. And instead of treating every form of redundancy, slack or variation as waste, we can ask whether some of it is precisely what makes the system robust.
