Imagine an institution in which the familiar behavioural causes of organisational fragmentation have largely disappeared. Departments share information when they are asked, specialists respect one another’s expertise, managers encourage collaboration and nobody deliberately hoards knowledge or protects professional territory. It might seem reasonable to expect such an organisation to have overcome the problem of silos, yet it could remain deeply fragmented for a reason that has little to do with unwillingness to cooperate. Its economists may describe problems differently from its engineers, its legal specialists may organise reality through distinctions that operational teams rarely use, its data analysts may rely on categories that correspond imperfectly with administrative responsibilities, and its frontline professionals may recognise patterns that cannot easily be represented in central reporting systems. Everyone can be competent and cooperative while the institution still struggles to make what they know work together.

The source of this difficulty lies in one of the great strengths of complex institutions: specialisation. Modern public organisations confront problems whose technical, legal, social and economic complexity exceeds what any generalist could reasonably understand in sufficient depth, so institutions divide cognitive labour among professions, departments and specialist units. Over time, this differentiation allows expertise to deepen. Specialists learn to perceive distinctions that outsiders overlook, accumulate experience within narrower domains, develop methods suited to particular classes of problems and create technical vocabularies capable of expressing forms of precision that ordinary language cannot easily sustain. In this sense, specialisation does more than divide tasks; it increases the cognitive resolution with which different parts of the institution can perceive reality.

That improvement is indispensable, but it has an architectural consequence. As expertise deepens, specialists do not merely learn more facts about a common object; they increasingly organise the object itself in different ways. A lawyer may understand a policy problem primarily through categories of authority, obligation and legal status, while an economist sees incentives, distributions and behavioural effects, a technologist sees data structures and system constraints, and a frontline professional sees recurring situations whose practical complexity fits awkwardly inside any of those abstractions. None of these perspectives has to be mistaken for integration to become difficult. They may all be accurate within their domains while highlighting different properties, using different categories and relying on different standards of relevance.

Specialisation therefore produces a paradox. The same process that makes the parts of an institution cognitively stronger can make the knowledge of the institution as a whole harder to integrate. Deeper local expertise generates differentiated languages, classification systems, information infrastructures, assumptions and standards of evidence, and those differences create what can be understood as epistemic boundaries: boundaries not merely between organisational units, but between different ways of knowing. As these cognitive structures become more sophisticated within each specialist domain, the work required to translate between them can grow even when everybody involved wants the translation to succeed.

Consider two competent teams trying to combine information about the same public problem. One organises its data around categories of citizens, while another organises its data around types of administrative process. Each classification may be entirely appropriate for the work that produced it, and both teams may share their datasets without hesitation, yet there may be no straightforward correspondence between the categories. Sending more information across the boundary does not necessarily solve the problem because the difficulty is not access but translation: what one system treats as the basic unit of analysis may not exist as such in the other.

This distinction matters because institutional fragmentation is often described through the language of silos, and the metaphor easily becomes a behavioural accusation. A silo is imagined as a group that refuses to collaborate, protects information or cares more about its own objectives than those of the wider institution. Those behaviours certainly occur, but treating them as the general explanation for fragmentation produces a misleadingly simple remedy: encourage people to collaborate more, share more information and break down cultural barriers. Such interventions can be useful when the problem is genuinely behavioural, but they may achieve surprisingly little when fragmentation has emerged from the structure of specialised cognition itself.

The difference becomes clearer if we return to the imagined institution in which everyone cooperates. Suppose every specialist answers requests promptly, meetings are constructive, data is shared freely and nobody regards expertise as private property. Even under these unusually favourable conditions, the specialists would still possess different bodies of knowledge, use different concepts, notice different aspects of situations and perhaps apply different standards for deciding what counts as persuasive evidence. Their information systems might still encode incompatible categories and their professional responsibilities might still lead them to frame the same problem differently. Perfect cooperation would therefore not produce perfect cognitive integration, because behavioural openness cannot by itself eliminate the epistemic boundaries created by differentiated expertise.

Recognising this distinction changes the diagnosis. Behavioural fragmentation occurs when people who could integrate knowledge fail or refuse to cooperate; structural cognitive fragmentation can persist even when they cooperate because the institution has differentiated its knowledge faster than it has developed the capacity to reconnect it. In the first case, incentives, trust and collaborative norms may be central. In the second, the institution faces a more architectural problem involving translation, interoperability, shared problem representation and the ability to expose differences between specialist assumptions without destroying the distinctions that make specialist knowledge valuable in the first place.

This is why the answer cannot simply be to reverse specialisation. Flattening professional distinctions into a common vocabulary would often eliminate precisely the cognitive depth that complex institutions require. An epidemiologist, an infrastructure engineer and a legal adviser should not be forced to understand every problem through identical categories merely because they work for the same government. The challenge is instead to recognise that every increase in useful cognitive differentiation creates a corresponding integration burden. Institutions need mechanisms through which different forms of expertise can meet, translate and sometimes contest one another while retaining enough of their internal structure to remain genuinely informative.

Artificial intelligence can intensify this paradox because it allows specialised functions to deepen their analytical capability without automatically improving institutional integration. A regulatory unit may use AI to develop more sophisticated risk classifications, a planning team may produce richer forecasts and a service organisation may identify patterns in cases that human analysts would struggle to detect. Each improvement can increase local cognitive capability, yet systems trained on different data, optimised for different objectives and organised around different classifications may generate outputs that become even harder to combine. Greater intelligence inside each component therefore does not guarantee greater intelligence at institutional scale; under some conditions, it may increase both cognitive power and cognitive fragmentation simultaneously.

This possibility reveals why institutional intelligence cannot be inferred from the sophistication of an organisation’s specialists, databases or AI systems considered separately. What matters is not only how much each component can know, but whether differentiated forms of knowledge can become usable together when a problem crosses the boundaries within which that knowledge was developed. The most difficult public problems routinely do exactly that, moving across legal, technological, social, economic and organisational domains whose specialist representations cannot simply be stacked beside one another and expected to form a coherent institutional view.

Specialisation remains one of the principal ways through which institutions become capable of understanding complex reality, and nothing in this argument diminishes its value. The deeper point is that cognitive depth has an architectural cost: as different parts of an institution learn to see more, they may increasingly learn to see differently. Specialisation can therefore make an institution smarter locally while making it more fragmented collectively, which means that institutional intelligence depends not on choosing between expertise and integration, but on building an organisation capable of preserving the depth of differentiated knowledge while continually creating the conditions under which those different ways of knowing can meet again.