The Intelligence Isn't in the Model. It's in the Architecture.
The question I hear most often from technology leaders entering serious AI adoption is some version of: how smart is the model? They want a capability benchmark, a rank, a number. It's the wrong question. The more useful question is: how well does your system distribute intelligence across all the nodes that do cognitive work?
Edwin Hutchins spent years studying how ships navigate. His 1995 book Cognition in the Wild made an argument that changed how cognitive scientists think about intelligence: when you watch a navigation team work, the intelligence isn't in any individual sailor's head. It's distributed across the navigator, the charts, the instruments, the radio, the institutional protocols, and the trained habits of the crew. Remove any node and the system degrades. The "thinking" happens at the system level, not the individual level. Hutchins called this distributed cognition.
Software delivery teams are distributed cognitive systems. A sprint planning session, a code review, a production incident: these aren't events where one smart person thinks and others execute. They're processes where reasoning is distributed across people with different knowledge, tools that encode prior decisions, documentation that carries institutional memory, and norms that structure how information flows. The team's collective intelligence is a function of the architecture, not just the talent.
AI extends this architecture. It adds a new class of node. And like every prior addition, what it does to system-level intelligence depends almost entirely on how it's integrated, not how capable the model is in isolation.
Hutchins' example is worth sitting with. A naval vessel has a navigation team, not a navigator. The team uses bearing timers, gyrocompasses, charts, depth finders, and published procedures. Each tool encodes knowledge that no individual needs to carry in their head. Each procedure ensures that information flows correctly between nodes. The intelligence of the system is greater than any individual's intelligence because the architecture amplifies and coordinates what each node contributes.
When GPS arrived, it didn't just give navigators better position information. It changed the distribution of cognitive work across the system. Some tasks moved from human cognition into the instrument. New failure modes appeared: what happens when the instrument is wrong and no one knows how to navigate without it? Distributed cognition doesn't just add capability. It shifts where reasoning happens and changes what breaks when a node fails.
AI is doing the same thing to knowledge work, faster and more extensively. The question isn't whether the model is capable enough to take on that role. The question is whether the system is designed well enough to support that node correctly.
An AI-native organization isn't one where AI does more work. It's one where the distributed cognitive architecture is designed with AI as a first-class node. Human judgment remains the highest-value node, but its role changes. It concentrates on novel situations, ethical tradeoffs, relationship context, and decisions where accountability matters. Model inference handles high-volume, well-defined reasoning tasks: summarization, classification, first-draft generation, pattern detection. The tooling and procedures are the modern equivalent of Hutchins' charts. Governance structures information flow the way navigation protocols do.
Most organizations approach AI as an add-on to existing work — a productivity tool that makes individual contributors faster. That produces AI deployments that improve individual task throughput without touching system-level intelligence. It's the equivalent of giving every sailor a faster calculator without changing how information flows between them, which charts they use, or which procedures govern their coordination. Individual nodes get faster. The system doesn't get smarter.
A 2025 paper examining AI integration in remote operations (air traffic control, industrial automation, intelligent ports) found that increased reliance on AI has profound implications not just for individual cognitive function but for team cognition. The finding wasn't that AI makes individuals better. It was that it changes the topology of how the team thinks, and that topology can be better or worse depending on how deliberately it was designed.
The more useful question for any AI program is not how capable the model is, but how well the system around it has been designed.