What is Collective Fluency?

Why individual AI proficiency doesn't automatically become organisational intelligence.

By Jordan Dalladay and Craig Wright | UN/DO

A new ventures team was given a mandate that is becoming increasingly familiar across organisations: use AI.

There wasn't much more precision than that. So people began experimenting in their own ways. One person used AI to investigate potential markets for an existing venture. Another used it to storyboard and prototype an idea still being validated. Others explored directional economics, marketing propositions and regulatory questions.

The tools were useful, but the most interesting part of the team's progress wasn't what happened between each person and the machine. It was what happened when they came back together.

Every week, they compared notes. People shared approaches that had worked, outputs that had looked convincing but proved unreliable, and situations where AI had helped them explore an idea more deeply than they otherwise could. They challenged one another's assumptions, borrowed useful techniques and brought what they had learned into the next week's work.

Over time, the team became remarkably effective at working with AI. But their advantage was not simply that several individuals had become proficient users.

They were already good at innovation work. They understood how to enter an ambiguous problem, identify what they didn't know, test assumptions against evidence and change direction when their understanding improved.

AI amplified an existing practice of inquiry, judgement and learning. Crucially, the team continued to develop that practice together.

This illustrates something we call Collective Fluency, one of the central ideas explored in The Last Mile of AI.

What is Collective Fluency?

Collective Fluency is the organisational ability for people with different levels and styles of AI capability to think, judge and learn together while individual practice continues to evolve.

It describes a capability that is different from the AI proficiency of any individual employee. Someone may be exceptionally skilled at prompting, evaluating models or building AI-assisted workflows, yet that expertise does not automatically improve the organisation's ability to reason and act collectively.

Individual fluency concerns how effectively a person works with AI. Collective Fluency concerns whether people can still develop shared understanding, exercise judgement and learn from one another as AI changes how each of them works.

This distinction matters because organisations don't create value merely by accumulating individual capabilities. They also depend on people interpreting situations together, reconciling different perspectives, challenging assumptions, agreeing what matters and coordinating action.

As AI becomes more capable, it can make individuals substantially more productive without necessarily strengthening those collective processes.

The organisation may have more intelligence available to it than ever before, while finding it increasingly difficult to understand how its people arrived at their conclusions.

Collective Fluency is the ability to prevent that gap from widening — and, potentially, to turn individual experimentation into a stronger shared capacity for thought and action.

The difference between using AI and thinking well with AI

Much of the conversation about AI fluency understandably focuses on the relationship between a person and a tool.

Can someone use the technology effectively? Do they know how to prompt it? Can they select an appropriate model, recognise obvious errors and produce something useful?

These are important skills. Working well with probabilistic systems involves judgement, experimentation and experience. People who spend significant time understanding the strengths and limitations of different models develop capabilities that newcomers do not yet possess.

But these are also the easiest capabilities to observe.

A presentation becomes more polished. A prototype that once took a week appears in an afternoon. Research is summarised more quickly, and someone produces an impressive analysis with relatively little visible effort.

From these outputs, we may reasonably infer that the individual has become more capable.

What we cannot necessarily infer is whether the underlying thinking improved.

A person might use AI to investigate an assumption more carefully, compare competing explanations or challenge a weak idea. Equally, they might use it to produce a more convincing version of an argument that should have been questioned in the first place.

The quality of the artefact does not always reveal the quality of the reasoning that produced it.

For an organisation, there is another question beneath that distinction: can other people still participate meaningfully in the reasoning?

An individual may arrive with a sophisticated recommendation built through many private interactions with AI. The resulting proposal might be coherent and well presented, but colleagues may have little visibility into which assumptions were tested, which alternatives were rejected or where uncertainty remains.

The individual has become faster. The organisation has not necessarily become better at thinking together.

This is why Collective Fluency cannot be measured simply by the number of employees using AI, their confidence with particular tools or the quality of the outputs they generate.

The deeper test is whether that growing capability remains available to collective judgement.

Not everyone needs to operate the machine

We encountered a useful example during an innovation sprint with a commercial strategy team working to improve a stubborn business metric.

The participants had very different levels of AI experience. Some used the tools daily. Others had never logged in.

One possible approach would have been to train everyone to a similar level before beginning the work. Another would have been to give each participant an AI workspace and let them explore independently.

Instead, AI was introduced at particular moments in a shared process of inquiry.

Early in the sprint, it helped bring customer personas, needs and journeys to life as teams developed potential interventions. Later, it supported the testing and refinement of ideas. As concepts became more concrete, it helped teams explore possible experiences and develop directional estimates of impact.

Each small team shared a single AI instance.

Often, the person most comfortable with the technology operated it. But the others were not passive observers. They supplied commercial context, challenged responses, questioned assumptions, connected emerging ideas to their understanding of customers and decided together what deserved further investigation.

One person might operate the machine, another might contextualise what it produced, and a third might recognise a contradiction that required the team to reconsider its approach.

The reasoning remained collective even though the technical capability was unevenly distributed.

This matters because the organisation does not necessarily need every person to develop the same proficiency with AI. An experienced commercial leader may contribute judgement that a highly proficient AI user lacks. A domain specialist may recognise when a plausible response contradicts operational reality. A less technically confident participant may ask the question that reveals an assumption everyone else has overlooked.

Different forms of expertise can remain valuable within one shared process of thought.

Capability can be distributed. Coherence has to be collective.

The objective is not to make everyone equally skilled at operating the technology. It is to design ways of working in which people can contribute meaningfully to the thinking, regardless of their technical fluency.

When AI turns teamwork into single-player mode

There is, however, an important counter-pattern.

In one technology team we encountered, AI became increasingly embedded in research, prototyping, planning, analysis, coding and testing. Individuals could move further independently, often completing work that previously required help from colleagues with different specialist skills.

The productivity opportunity was genuine. Fewer tasks needed to wait for scarce expertise, and people could explore and develop ideas at a pace that would previously have been difficult.

But new problems began to emerge.

Features could be developed in isolation before the wider team had interrogated their assumptions. Plans arrived looking complete while important context remained with the person who had generated them. Testing practices became uneven, and lessons from failures were not always discussed or carried into subsequent work.

The organisation had reduced some of the friction of collaboration. It had also reduced some of the occasions on which shared understanding used to develop.

This is a subtle consequence of increasing individual capability.

When a person needs help from colleagues to develop an idea, the work often involves explanation, disagreement and negotiation. People discover that they have understood the problem differently. An assumption is challenged, a requirement changes or someone reveals a constraint that was not visible from the original perspective.

Some of those interactions are unnecessary overhead, and AI can usefully remove them. But others are part of how the organisation thinks.

If AI enables people to avoid waiting for colleagues without preserving the processes through which shared meaning develops, individual productivity can increase while collective coherence weakens.

The issue is not simply that people communicate less. It is that they may communicate later, after more work has already been built around assumptions that were never collectively examined.

A polished proposal can be harder to challenge than an unfinished idea. A partially implemented feature can create greater resistance to reconsideration than an early conversation about what should be built.

AI may not remove the need for collective judgement. It may postpone the point at which that judgement happens.

And postponed judgement can become more expensive.

Polished output is not shared understanding

AI makes it easier to create artefacts that look complete.

Strategies are clearly structured. Research reports appear comprehensive. Requirements documents are detailed, and presentations communicate a compelling argument.

These improvements can be valuable. Better documentation and clearer communication can help organisations work more effectively.

But a shared artefact is not necessarily evidence of shared understanding.

A team may approve a polished document while its members hold different interpretations of the problem, the assumptions behind the recommendation or the meaning of success. The language appears coherent because the document is well written, not because the organisation has developed a coherent perspective.

We distinguish this in The Last Mile of AI as the difference between artefact quality and conceptual cohesion.

Artefact quality concerns how well an output is expressed. Conceptual cohesion concerns whether people have developed enough shared meaning around it to coordinate, challenge and act effectively.

The two can reinforce one another, but they are not interchangeable.

As AI improves artefact quality, the distinction becomes more important. A convincing output may conceal the fact that critical questions have not been discussed or that the people responsible for acting upon a recommendation do not share the same understanding of why it matters.

Collective Fluency requires organisations to pay attention not only to what has been produced, but to whether the reasoning behind it has become sufficiently intelligible to others.

Sometimes that requires more explicit discussion. Sometimes it requires better ways of exposing assumptions, uncertainty and competing interpretations. Sometimes it means involving the people responsible for the consequences earlier, before a recommendation has become difficult to reconsider.

The objective is not perfect agreement. Productive disagreement may be precisely what improves the thinking.

The objective is enough shared understanding that disagreement, judgement and action remain possible.

The most useful practice may not be the prompt

As people become more experienced with AI, they naturally develop techniques worth sharing.

An engineer discovers an effective way to evaluate AI-assisted code. A strategist develops a useful method for testing a market hypothesis. A salesperson learns how to prepare for difficult customer conversations, while a researcher finds a way to use AI to compare alternative interpretations of evidence.

The obvious organisational response is to capture these techniques.

Save the prompt. Document the workflow. Add it to a shared library. Turn it into training so other people can reproduce the result.

Sometimes that works well. Reusable prompts, templates and processes can reduce unnecessary experimentation and help effective practice spread.

But the most valuable knowledge may not be contained in the visible technique.

It may sit in how the person framed the uncertainty, what they were looking for in the response, which assumptions they knew to challenge and how they recognised when the output was no longer trustworthy.

The same prompt may behave very differently in another person's hands because that person lacks the context or experience that made the original practice effective.

Capturing the artefact is not always the same as transferring the judgement.

This is why Collective Fluency requires organisations to make more of the reasoning around successful practice discussable.

Rather than asking only which prompt someone used, we might ask what they were trying to establish, why they chose that approach, what they delegated to the model, what they checked independently and where they knew not to trust the result.

Those questions make the underlying practice more transferable without assuming that everyone should follow one identical workflow.

They also allow people to learn from failure. An output that looked convincing but proved wrong may teach the organisation more about effective judgement than another example of a successful prompt.

In the ventures team, this was a central part of the weekly conversation. People were not simply exchanging instructions for the machine. They were making their judgement visible to one another.

Private technique was becoming collective practice.

Collective Fluency is not uniformity

It would be easy to turn this argument into another programme of standardisation.

Choose approved tools, establish best-practice prompts, create a common training curriculum and require everyone to follow the same process.

Some standardisation is necessary. Organisations need appropriate safeguards, common expectations about evidence and security, and enough consistency that work can be understood and coordinated.

But uniform practice is not the same as Collective Fluency.

Different disciplines face different problems and require different forms of judgement. An engineer testing generated code, a finance leader evaluating an investment case and a product designer exploring an emerging customer need should not necessarily interact with AI in the same way.

Even within one team, people may develop different levels of proficiency and different approaches to the same task.

Trying to eliminate those differences could suppress useful experimentation. Ignoring them entirely, however, can make individual practice so private and specialised that colleagues can no longer understand or challenge what is happening.

The useful middle is shared intelligibility rather than identical technique.

People should be able to explain what role AI played in their work, what evidence they are relying on, what remains uncertain and which decisions still require human judgement.

Some techniques should spread widely. Others should remain local because they depend on particular expertise, circumstances or problems.

The aim is not to make everyone work the same way. It is to ensure that different ways of working remain capable of contributing to a coherent organisational process.

From individual learning to organisational capability

There is a broader organisational learning problem here.

An individual can become exceptionally skilled through years of experimentation. But if that experience remains private, the organisation may benefit only while that person is present and working on the relevant problem.

A team may learn something important during a project, yet struggle to make that learning available to others. A successful AI-assisted approach may be repeated without anyone understanding why it worked, while an earlier failure is rediscovered elsewhere because the reasoning behind it was never shared.

These are familiar organisational challenges. AI did not create them.

What AI changes is the speed and variety of experience being generated. People can undertake more experiments, consider more alternatives and produce more work than before. The opportunity for learning expands, but so does the possibility that valuable experience remains fragmented.

Collective Fluency helps connect individual experimentation with organisational learning.

That might mean creating opportunities for teams to compare approaches, making critical assumptions explicit, recording why a decision was made or adapting a workflow when a recurring problem becomes visible.

Not every useful lesson needs to become a permanent rule. Some practices depend on context, and some lessons will become obsolete as the technology changes. Preserving the capacity for judgement matters as much as preserving the results of previous judgement.

The important question is whether learning changes what the organisation is capable of doing next.

A collection of prompt libraries may improve access to past techniques. A collectively fluent organisation is better able to understand which techniques still make sense, when to adapt them and when to question the assumptions behind them.

That is a more durable capability.

How can organisations develop Collective Fluency?

Developing Collective Fluency does not necessarily require another enterprise-wide training programme. It begins with examining how people actually use AI in the work and whether the organisation's existing arrangements help or hinder shared reasoning.

A useful starting point is to make experimentation discussable. Teams need opportunities to explain where AI helped, where it failed, which aspects of the output they trusted and which required independent judgement. This should not become a ritual of showing impressive outputs or celebrating adoption. The value lies in making the underlying decisions and uncertainties available for discussion.

The next consideration is where collective thought should happen. Some work benefits from independent AI-assisted exploration. Other work requires early challenge, cross-functional interpretation or agreement about what problem is being solved. Organisations should be deliberate about those moments rather than assuming every activity either needs a meeting or can be completed independently.

It is also important to distinguish the different roles people play within AI-enabled work. One person may be technically proficient, another may supply essential domain context, and another may have responsibility for the decision or its consequences. Effective collaboration does not require identical contributions, but it does require clarity about how those contributions interact.

Over time, organisations can examine whether useful learning is travelling beyond the individuals who discovered it. Are techniques being copied without their underlying rationale? Can people explain the assumptions behind AI-assisted recommendations? Do people with different levels of experience remain able to challenge one another's conclusions? When the team learns that an approach is unreliable, does future work actually change?

None of these questions needs to become another universal maturity score. They are ways of investigating whether growing individual capability is strengthening the wider organisation.

This is also where Collective Fluency connects with Cognitive Infrastructure: the organisational conditions through which people and machines frame problems, exercise judgement, coordinate action and learn.

Collective Fluency is one of those conditions. It describes how an organisation preserves and develops shared reasoning while individual capabilities become more varied and more powerful.

Why Collective Fluency matters to the Last Mile of AI

The Last Mile of AI concerns the organisational challenge of turning available intelligence into meaningful outcomes.

As AI capability improves, the effort required to generate analysis, ideas, recommendations and artefacts can fall dramatically. Yet those outputs still need to be interpreted, challenged, coordinated and translated into action.

Collective Fluency matters because much of that work is inherently social.

People need enough shared understanding to decide which problems matter, whether evidence is credible, what trade-offs are acceptable and how different contributions fit together.

When individual AI capabilities grow faster than the organisation's ability to maintain that understanding, the constraint on progress can move. What was once limited by production capacity may become limited by coordination, judgement or the team's ability to learn together.

This is one form of Constraint Migration.

It can also help explain why improvements in individual task productivity do not always translate into comparable gains in Outcome Throughput. Producing work faster is valuable, but if colleagues cannot understand, evaluate or integrate it, the wider system may not progress at the same rate.

The economic consequences are visible through Intelonomics as well. Cheaper intelligence production may create new demands for interpretation, review and shared judgement. Collective Fluency can help organisations manage those demands more effectively, not simply by producing less work, but by making the reasoning behind it more accessible and useful.

These are different perspectives on the same broader organisational challenge: ensuring that intelligence becomes consequential rather than merely abundant.

From more fluent individuals to a more intelligent organisation

AI is expanding what individuals can do. People can undertake work that previously required specialist support, explore unfamiliar domains and develop ideas at speeds that would once have seemed extraordinary.

That is a genuine opportunity.

But an organisation is not simply a collection of independently capable people. Its effectiveness also depends on shared concepts, productive challenge, coordinated decisions and the ability to learn from experience.

As AI enables individuals to work in increasingly different ways, preserving those collective capabilities becomes more important, not less.

The challenge is not to ensure everybody uses the same technology in the same way. It is to create the conditions in which different capabilities can remain part of one coherent process of thought.

Collective Fluency is what allows individual learning to contribute to shared judgement, and shared judgement to improve what the organisation does next.

An organisation does not become collectively fluent simply because everyone can talk to a machine. It becomes fluent when working with the machine helps people think, judge and act better — and still allows them to think together.

Explore The Last Mile of AI

The Last Mile of AI: What happens when intelligence is no longer scarce? is a forthcoming book by Jordan Dalladay and Craig Wright, co-founders of UN/DO, a strategic futures studio.

The book explores how the changing economics of intelligence affect organisational performance, why constraints shift as AI capability improves, and what leaders need to rethink to turn technological possibility into meaningful progress.

Coming December 2026.

Discover the book and explore the wider Last Mile series

Explore the Concepts Behind the Last Mile of AI

The Last Mile of AI is more than a single challenge. Explore the interconnected ideas that help explain why intelligence doesn't automatically become progress - and what organisations can do about it.

Key Concepts

Collective Fluency - How people with different levels of AI capability can continue to think, judge and learn together - turning individual proficiency into collective intelligence.

Constraint Migration - Why improving one capability often reveals another bottleneck, shifting what limits organisational progress.

Intelonomics - The changing economics of intelligence - from the cost of producing it to the effort of absorbing it and the consequences of acting upon it.

Outcome Throughtput - Why making individual tasks faster doesn't necessarily mean making more progress - and how to measure the outcomes that matter.