OKRs work brilliantly on complicated systems and fail on complex ones
Most organisations adopting OKRs today are using a tool built for engineering teams to solve problems that are not engineering problems. The framework works. It often works very well. It does not work everywhere.
I spent the first part of my career as an internal auditor, and one of the questions auditors learn to ask is whether a control is actually controlling. Apply the same question to OKRs and the answer becomes uncomfortable. In the organisations where OKRs travel furthest, on engineering teams, on sales pipelines, on logistics operations, the framework controls something real. In the organisations where leaders are reaching for OKRs because their environment feels out of control, the framework controls almost nothing. It manufactures an appearance of strategic discipline while the actual system the organisation is meant to influence keeps moving in directions the OKR document did not anticipate.
The interesting and uncomfortable thing about this is that the organisations most desperate to adopt OKRs are exactly the organisations where OKRs are structurally most likely to fail.
Two kinds of system
To make this argument I need to make a distinction that, once seen, is hard to unsee. There are systems that are complicated, and there are systems that are complex. The two words sound interchangeable in everyday speech. In the systems literature they are precise terms with very different implications.
A complicated system is one with many moving parts that interact through legible cause-and-effect chains. A jet engine is complicated. A logistics network is complicated. A modern software platform is complicated. The number of components is enormous, the interactions are not trivial, but expert understanding is possible. Given enough time, a sufficiently skilled team can take the system apart, understand each piece, predict how a change in one place will affect another, and reassemble it. Complicated systems are knowable. They reward planning. They reward the kind of structured execution that OKRs were designed for.
A complex system is something else. As John Miller and Scott Page describe in Complex Adaptive Systems, a complex system is one whose behaviour emerges from the interactions of its parts in ways that cannot be predicted from the parts themselves. A national economy is complex. A city's housing market is complex. A public's relationship with a fast-changing technology is complex. You can study the components forever and still be surprised by what the system does. Goals set against the system's current behaviour become obsolete as the system adapts to having those goals pursued at it.
This distinction matters because the most popular management frameworks of the last twenty years, OKRs prominent among them, were designed in the world of complicated systems. They assume that if you set a clear objective, choose three or four measurable key results, and review them on a quarterly cadence, the gap between intent and outcome is mostly an execution gap that the framework will close. In a complicated system, that assumption is true often enough that the framework looks like magic. In a complex system, the assumption is the source of the failure.
What OKRs are actually good at
I want to be fair to John Doerr's framework, because most critiques of OKRs collapse into critiques of how OKRs are used badly, and that is not the argument I am making.
Measure What Matters is, on its own terms, a strong book. Doerr describes a tool that has worked extraordinarily well at Intel, at Google, at a number of foundations and high-growth companies. The OKR mechanic, set an ambitious objective, attach a small number of measurable key results, score quarterly, distinguish committed from aspirational, has genuine virtues. It forces clarity. It exposes drift early. It makes priorities legible across a team. In environments where the work is largely a question of execution against a known target, OKRs do exactly what they promise.
Look closely at the canonical OKR success stories and a pattern appears. Intel's microprocessor business in the 1980s. Google's search and advertising infrastructure in the 2000s. Engineering organisations shipping software. Sales organisations working a defined pipeline. Logistics teams optimising throughput. These are not random domains. They are domains where the underlying system is complicated rather than complex. The cause-and-effect chains are reasonably stable. The goal is roughly knowable a quarter ahead. The team can plan, execute, and measure, and the system will mostly cooperate.
This is not a small set of organisations. There is a great deal of important work in the world that fits this description, and OKRs are an appropriate tool for it. My quarrel is not with the framework. It is with the assumption that the framework generalises.
Why complex systems break the cycle
The OKR cycle has three implicit assumptions. The objective stays relevant for the period. The key results are reasonable proxies for progress on the objective. The team can move the key results through its own action. In a complicated system, all three usually hold. In a complex system, any of them can fail mid-cycle.
Miller and Page's central claim, drawn from decades of computational modelling, is that complex systems generate emergent behaviour. The system does not respond to your intervention in proportion to your intervention. It responds in ways that depend on what the rest of the system is doing at the same time. Set a goal about adoption, and the population whose adoption you are measuring adapts to your campaign. Set a goal about reducing a behaviour, and the behaviour reorganises itself around your measurement. The famous problems of perverse incentives in policy, from Soviet nail factories producing useless oversized nails to British surgeons declining the riskiest patients to keep their published mortality rates clean, are not stories about bad people. They are stories about complex systems reacting to fixed goals in ways the goal setter did not anticipate.
David Peter Stroh, in Systems Thinking for Social Change, develops this for the field he knows best, the work of NGOs and public agencies trying to shift social outcomes. His argument is that interventions in complex social systems routinely produce unintended consequences because the system's feedback loops are richer than the planning model accounts for. A homelessness programme that hits its housing placement target can simultaneously reduce its capacity to address the upstream pressures producing homelessness in the first place. A school funding formula that hits its equity target can entrench the very segregation it was designed to undo. The goal was met. The system shifted underneath the goal. The next quarterly review will report green status on a problem that has quietly worsened.
Ben Ramalingam, in Aid on the Edge of Chaos, documents the same pattern across decades of international development. The aid sector, faced with complex adaptive challenges in fragile states, has consistently reached for planning instruments that assume the world stands still while the intervention runs. Logical frameworks. Theories of change. Quarterly reporting against pre-set indicators. These are the cousins of OKRs in a different vocabulary. Ramalingam's argument, supported by case after case, is that they routinely produce the appearance of progress while the underlying situation either fails to improve or actively worsens, because the planning instrument cannot keep up with what the system is doing.
The pattern is the same in all three accounts. The framework assumes a stable target. The system makes the target unstable. The framework keeps reporting against the original target. The gap between what is being measured and what is happening widens, slowly at first, then visibly, then catastrophically.
The organisations most exposed
This is where the non-obvious insight lives. The organisations most attracted to OKRs are not, on the whole, engineering teams that already know what they are doing. They are organisations that feel out of control and are looking for a framework that promises to restore strategic discipline. That category, almost by definition, is dominated by organisations operating in complex environments rather than complicated ones.
Take a large retail bank deciding, somewhere around 2024, that it needs to set OKRs for its AI adoption programme. The objective is reasonable, capability uplift across the workforce, integration of large language models into a handful of customer-facing services, responsible deployment throughout. The key results get drafted. So many employees trained by Q2. So many use cases in production by Q3. So many customer interactions AI-assisted by year end. Then a major lab releases a new model class in March that changes what is even possible and makes the chosen vendor's offering look suddenly weak. A regulator publishes guidance in May that ringfences which decisioning use cases the bank can responsibly deploy AI into. A core platform vendor's roadmap shifts in June and the integration the Q3 target depends on disappears. Meanwhile the workforce is adapting to the tooling at very different speeds, with some teams discovering uses no one anticipated and other teams quietly avoiding the rollout altogether. By the September review the OKRs no longer describe the work that is actually happening. The team is, in effect, choosing between honestly reporting that the OKRs are obsolete and quietly re-baselining to make the dashboard green. Neither response captures what the organisation actually learned that quarter.
Take a regional education department rolling out AI-assisted personalised learning across its primary schools, with the genuine intent of improving foundational literacy for the children who are furthest behind. The objective is honourable. The key results get attached, tools deployed in so many schools by Q2, teachers trained by Q3, measurable improvement in standardised reading scores by year end, equity gap narrowed by a defined margin. None of these are unreasonable on paper. All of them sit inside a complex social system. Teachers adapt to the tool unevenly, some integrating it thoughtfully into their practice, others abandoning it after a fortnight. The students who benefit most turn out to be the ones the intervention was not primarily designed for, because the children who were furthest behind also have the least support at home to use the tool consistently. A privacy controversy in another country in April changes the local political context, parental opt-outs spike, and two school boards request that deployment be paused. A new ministry guidance document in June restricts the data that can be used for personalisation, which forces a redesign of the very feature the equity case rested on. By the September review the system has learned an enormous amount, almost none of which is captured by the original key results. The team is left choosing between reporting honest obsolescence or quietly re-framing the targets so the year-end report can stay positive. The children the intervention was supposed to serve are, by then, no closer to the literacy outcome that motivated the programme in the first place.
Take a humanitarian agency running cross-country programmes in protracted crises. The objective is to deliver a defined level of service to a target population across multiple countries. The key results are quantitative, beneficiaries reached, services delivered, partnerships activated. Two of the countries see a sudden political shift that closes humanitarian access. A third sees an unforeseen influx of displaced people that triples the need overnight. A fourth's currency collapses and the budget no longer buys what the plan assumed. The OKR framework gives the agency no good way to account for any of this within the cycle, and the temptation to either claim partial credit against the original numbers or quietly rewrite history is structural rather than personal.
In each case, the people involved are not stupid and they are not lazy. They are trying to do serious work inside a framework that assumes their problem is simpler than it is. The pattern is also visibly sector-agnostic. The bank, the education department, and the humanitarian agency look very different from the outside. The failure mode is the same. The framework launders strategic incoherence into apparently rigorous quarterly reporting, until enough cycles accumulate that the gap between the documents and the reality becomes undeniable. By then, the cost is years of misallocated effort and a leadership team that has lost the ability to distinguish progress from theatre.
What to do instead is not another framework
The temptation, when arguments like this one are made, is to ask for the replacement framework. There is no replacement framework, at least not in the sense people usually mean. What complex environments require is not a different set of templates. It is a different posture toward goals themselves.
The posture has three elements that I think are non-negotiable. The first is that the unit of commitment is directional intent, not numeric target. Leaders in a complex environment commit to where they are trying to take the organisation and why, not to the precise indicator that will move by what amount by what date. The numeric targets are tools for orientation, used loosely, replaced often. They are never the contract.
The second is that the feedback loop has to be much shorter than the planning loop. In a complicated system, you can plan a quarter and execute against it. In a complex system, you have to be willing to learn faster than you plan. That means weekly, sometimes daily, conversations about whether the system is doing what you expected and what to change if it is not. It also means being willing to invalidate last week's commitment in this week's review, without that being treated as a failure of will.
The third, and the hardest, is that revising the goal has to be normal. In an OKR culture, changing the objective mid-cycle is a sign of poor planning. In a complex environment, refusing to change the objective when the system has clearly shifted is a sign of poor leadership. The discipline being asked for is the discipline of being seen to change your mind in public, repeatedly, without that being read as evidence that you did not know what you were doing in the first place.
That is a much harder thing to install than an OKR template. It requires a culture that can distinguish drift, where the organisation loses its sense of direction, from adaptation, where the organisation revises its direction because the world has changed. Most management cultures cannot tell the two apart, which is part of why frameworks built for complicated systems remain so attractive even where they do not fit.
Think about it.
Sources and methods
Doerr, John. Measure What Matters: How Google, Bono, and the Gates Foundation Rock the World with OKRs.Portfolio, 2018. The canonical statement of the OKR framework and the case studies that underpin its popularity.
Miller, John H., and Scott E. Page. Complex Adaptive Systems: An Introduction to Computational Models of Social Life. Princeton University Press, 2007. The theoretical grounding for the distinction between complicated and complex systems used in this article.
Stroh, David Peter. Systems Thinking for Social Change: A Practical Guide to Solving Complex Problems, Avoiding Unintended Consequences, and Achieving Lasting Results. Chelsea Green Publishing, 2015. The practitioner's case for systems thinking in social interventions, including extensive treatment of perverse outcomes from goal-driven programmes.
Ramalingam, Ben. Aid on the Edge of Chaos: Rethinking International Cooperation in a Complex World. Oxford University Press, 2013. A career synthesis of how complex-systems thinking should reshape the planning and accountability instruments of international development.
