The intelligent organisation is not the one with the most AI

There is a familiar story about artificial intelligence in organisations. It says that intelligence arrives through software: deploy enough models, automate enough workflows, give every employee a copilot, and the organisation becomes smarter.

That story is appealing because it makes transformation sound like a procurement exercise. But it confuses computational capability with organisational intelligence.

An organisation is intelligent when it can notice what matters, bring the right knowledge to bear, make sound decisions under uncertainty, learn from experience and adapt. AI can contribute to each of those capacities. It cannot substitute for the institutional conditions that make them possible.

That is the useful provocation in Vegard Kolbjørnsrud’s article, “Designing the Intelligent Organization: Six Principles for Human-AI Collaboration.” The article defines organisational intelligence as an organisation’s ability to acquire and apply knowledge to solve problems and adapt. Its central point is simple but easily missed: using AI to replace people may make an organisation more efficient, without necessarily making it more intelligent.

At a moment when governments, universities, businesses and civil-society organisations are all being urged to “adopt AI”, that distinction deserves much more attention.

We have not agreed on what intelligence is

There is a prior problem in all this enthusiasm: there is no commonly held definition of intelligence. Psychologists, philosophers, computer scientists and educators have long disagreed about whether intelligence is primarily reasoning, learning, problem-solving, adaptability, social understanding, creativity, consciousness, or some combination of these things.

That disagreement does not prevent useful work. But it should make us cautious about casually describing a system as intelligent simply because it performs impressively on a benchmark or completes a task that used to require human labour.

Large language models have made this problem much harder to ignore. They can produce fluent prose, sustain a plausible conversation, summarise complex material and adopt the linguistic mannerisms of expertise. They sound smart. But sounding intelligent is not the same as understanding, exercising judgement, possessing knowledge in the human sense, or being accountable for the consequences of an answer.

This is not an argument that LLMs are useless or that their capabilities are trivial. It is an argument against allowing linguistic fluency to settle questions that are organisational, ethical and political. A confident response can conceal fragile reasoning, missing context, inherited bias or complete fabrication. The more persuasive the language, the more important it is to retain the capacity to test, question and contextualise it.

Kolbjørnsrud’s account is helpful precisely because it moves the focus away from the metaphysical question of whether a machine is truly intelligent. The practical question is whether a collective of people, digital systems, structures and norms can acquire and apply knowledge to solve problems and adapt. That is a much more demanding standard than producing a convincing paragraph.

Efficiency is not the same as intelligence

Automation can be valuable. It can remove drudgery, increase speed, reduce error and make services available at a scale that would otherwise be impossible. Those are real gains. They are not, however, proof that an organisation has become better at understanding the world or acting well within it.

Consider a public agency that uses an AI system to triage routine applications. If the system reduces processing time, the agency has gained efficiency. But if the agency cannot identify when the system is unsuitable, cannot explain a contested decision, and has no meaningful pathway for staff or citizens to challenge errors, it has not gained much intelligence. It may simply have made its existing blind spots faster.

The same is true in corporate settings. A generative-AI tool that produces reports in minutes is helpful. Yet the central questions remain: are people asking better questions, testing the output against reality, seeing what the model cannot see, and changing course when evidence requires it?

The question for leaders is therefore not, “Where can we insert AI?” It is, “Which human and machine capabilities do we need to combine to solve the problems we actually have?”

Design for complementarity, not imitation

One of the paper’s six principles is relevance: the type of intelligence must match the nature of the problem. This matters because human and machine capabilities are different, uneven and context-dependent.

Machines are often excellent at analysing very large volumes of data, recognising patterns and generating probabilistic predictions. Humans remain essential where the task is ambiguous, values are contested, the evidence is incomplete or the consequences of a decision are distributed unevenly across people and communities. These are not marginal cases. They are much of what matters in public policy, healthcare, education, safety, justice, leadership and governance.

The strongest human-AI arrangements therefore do not try to make machines look human, or humans behave like predictable components in an automated system. They distribute work according to complementary strengths.

That framing changes the way we think about AI capability. An AI system that is unlike us can be more useful than one that merely mimics us. A model that finds patterns in infrastructure data, for example, may complement engineers’ contextual knowledge, operational experience and responsibility for public safety. It should not be mistaken for a replacement for that judgement.

This is also why the most consequential design decisions are rarely technical alone. They concern who gets to define the problem, which data count as evidence, what outcomes are optimised, who can intervene, and whose knowledge is treated as relevant.

Do not automate people into irrelevance

Kolbjørnsrud makes a distinction that should be pinned to the wall of every AI transformation office: replacing intelligent humans with intelligent machines does not automatically make an organisation more intelligent. It may simply make it more efficient.

The more interesting possibility is what happens after routine work is automated. Does the organisation redeploy people into work that makes fuller use of their expertise? Do case workers have more time for difficult cases? Do researchers have more time to investigate, interpret and communicate? Do managers have more capacity to listen, deliberate and make accountable decisions?

Or does automation simply intensify work, narrow discretion and reduce the number of people able to understand how a service actually operates?

The answer has implications for workforce strategy and for AI governance. Reskilling cannot mean teaching a small technical team to operate new tools while everyone else is expected to accommodate them. It must include building broad AI literacy, but also preserving and developing the human capabilities that automated systems cannot supply: critical thinking, ethical reasoning, domain expertise, empathy, contextual awareness and the ability to recognise a category mistake before it causes harm.

In other words, an intelligent organisation should use AI to stop people doing machine work, not to turn people into peripheral supervisors of machines.

Diversity is an intelligence strategy

The paper’s diversity principle is especially important in an era of increasingly standardised AI systems. Complex problems are better approached by groups with genuinely different knowledge, skills and perspectives. Homogeneous teams can be fast, harmonious and catastrophically wrong.

AI can expand a team’s cognitive diversity when it surfaces patterns, generates alternatives or makes dispersed information usable. But AI can also reduce diversity if every team relies on the same model, trained on similar material, operating through the same assumptions and prompts.

This is a governance issue. “Human in the loop” is not a meaningful safeguard if the human is merely there to ratify the system’s recommendation. Nor is consultation meaningful if the people consulted share the same professional background, institutional incentives and worldview as the people who built the system.

Real diversity introduces friction. It can slow a decision, expose disagreement and make an apparently neat solution more complicated. That is not necessarily a failure of collaboration. It is often the point.

For organisations working in high-stakes settings, the relevant question is not whether an AI system has been reviewed by a human. It is whether the people shaping and supervising it bring sufficiently varied forms of knowledge to identify harms, challenge assumptions and imagine consequences beyond the immediate task.

Collaboration needs architecture, not slogans

“Collaboration” is one of those words that can lose all meaning through repetition. Yet human-AI collaboration is not automatic. It requires deliberate organisational design.

People need to know what a system can do, where it is likely to fail, how to question it and when to disregard it. They also need practical routes to share what they learn: escalation paths, feedback loops, documented decisions, usable incident reporting, opportunities to improve workflows, and time to reflect on failures.

This is where culture and architecture meet. A psychologically safe culture is not enough if people have no authority to intervene. A sophisticated governance policy is not enough if frontline workers cannot understand or use it. A central AI team may develop standards and reusable infrastructure, but intelligence will remain concentrated unless the rest of the organisation can participate in shaping how AI is used.

We should be wary of organisations that treat AI governance as a control function located at the centre, while treating experimentation as something done at the edge. Good governance must support distributed learning. The people closest to a service, a customer, a community or an operational risk often see what senior leadership and model developers cannot.

The work is less about building a hierarchy of human overseers and more about building reliable relationships among people, systems, evidence and decision rights.

Explanation is a democratic capability

The sixth principle in the paper is explanation: intelligent organisations seek explanations and act responsibly. This is not just a technical preference for interpretable models. It is a question of power.

When consequential decisions are delegated to systems that cannot be questioned, institutions accumulate what the article calls “intellectual debt”. They may get an answer without knowing why it is reliable, when it is applicable, what it leaves out, or how to correct it when circumstances change.

Explanations matter differently to different people. A data scientist may need to understand a model’s behaviour. A clinician may need a meaningful account of why a recommendation should influence care. A member of the public may need to know the basis on which a decision affecting them was made and how they can challenge it. Governance that offers only a technical explanation to technical staff is not explanation enough.

This is why responsible AI cannot be reduced to compliance checklists. Legality matters, but it does not settle questions of legitimacy, fairness, environmental impact, social trust or democratic accountability. Those are matters for human judgement, and judgement requires institutions willing to surface trade-offs rather than hide them behind the authority of a model.

Move fast, and responsibly

There is a great deal of pressure to deploy AI quickly. Some of it is legitimate: technological change is rapid, competitive positions can shift, and public expectations are rising. But the old technology mantra of moving fast and breaking things was always an odd fit for institutions responsible for people’s livelihoods, health, safety, rights and essential services.

The alternative offered by the paper is more useful: move fast and responsibly.

That does not mean waiting for perfect certainty. It means treating speed as one objective among several, alongside contestability, safety, inclusion, sustainability and public trust. It means piloting in ways that create evidence, not merely publicity. It means being prepared to pause, change direction or withdraw a system when its harms outweigh its benefits.

Most of all, it means recognising that AI governance is not a brake on intelligence. It is part of the organisational capability to learn, correct itself and adapt.

The organisations most likely to thrive in the AI era will not be those that automate the fastest. They will be the ones that make better use of human judgement, welcome genuinely different perspectives, build systems people can understand and challenge, and remain accountable for the choices they make.

That is what an intelligent organisation looks like.