By Jordan Dalladay and Craig Wright | UN/DO
A small software company made a deliberate decision to become more AI-native. The ambition was to use artificial intelligence wherever it could remove friction from the work, from research and planning to prototyping, coding and communication. People were encouraged to think less like individual producers and more like orchestrators of multiple AI capabilities.
And it worked. Tasks that previously required specialist support could be undertaken independently, ideas could become working prototypes much faster, and plans and code could be produced in a fraction of the time. The organisation had removed some of the constraints that previously limited how quickly people could work.
But as the team examined its new ways of working, a different set of problems started to emerge.
AI-assisted documents were becoming large enough that reviewing them properly was difficult. Features could be developed faster than they could be adequately tested, and work produced independently was sometimes difficult to reconcile with the wider product. More plans were being generated than the team could meaningfully evaluate, while lessons from failed experiments were not always finding their way into shared organisational understanding.
The technology hadn't failed. Quite the opposite: it had succeeded in making production faster, exposing weaknesses in the surrounding system that had previously been less consequential.
This is an example of constraint migration, one of the central ideas explored in our book, The Last Mile of AI.
Constraint migration occurs when improving one part of a system makes a different limiting condition more consequential. In AI transformation, it describes how successful improvements to technological or individual capabilities can shift the effective bottleneck towards other parts of the organisation, including judgement, attention, coordination, authority, incentives and learning.
The concept builds on a familiar principle from systems thinking and constraint-based management: the performance of a system is shaped by its limiting conditions, not simply by the capabilities of its individual components.
A manufacturing line does not necessarily produce twice as much because one machine doubles its speed. If the next machine cannot process the additional work, the queue simply moves. Improving the original bottleneck may be valuable, but it changes which part of the system needs attention next.
Knowledge work operates according to similar principles, although its constraints are often less visible. Rather than physical inventory accumulating between machines, work may accumulate as proposals awaiting decisions, analysis requiring interpretation, outputs needing verification or teams waiting for agreement about what happens next.
AI makes this particularly important because it can transform the capacity of individual activities much faster than the surrounding organisation changes.
When research, writing, analysis, design or software development becomes substantially faster, other capabilities may become the effective limits on progress. Those capabilities may not have deteriorated. Their relative importance has changed because the work reaching them has changed.
The important question is no longer simply what AI has made faster, but what now prevents the organisation from progressing.
We tend to approach transformation as though the problem being solved will remain stable throughout the intervention. An organisation identifies a bottleneck, invests in a capability to address it and expects the resulting improvement to translate into greater performance.
Sometimes that is exactly what happens. A repetitive task disappears, a costly process becomes cheaper or a previously limiting activity can finally keep pace with demand. The improvement may flow through to the outcome without requiring substantial changes elsewhere.
But organisations are systems of interdependent capabilities, and the relationship between those capabilities changes when one of them improves.
Returning to the software company, the original constraint was partly the effort involved in producing work. Creating a prototype required time, writing substantial code required technical capacity, and developing a detailed proposal involved enough effort to limit how many could be pursued.
Those limits were frustrating, but they also influenced the organisation's way of working. When developing a proposition required several people, context was exchanged during the process. When building something was expensive, teams had reasons to discuss what deserved to be built. When work took longer to produce, there were natural limits on how much could arrive for review.
None of this means that slowness was inherently valuable. Much of the friction was unnecessary, and reducing it created genuine opportunities. But the previous constraints had also shaped how people coordinated their work, sometimes providing functions the organisation had never consciously designed.
Once AI reduced the cost of production, some of those interactions no longer happened automatically. People could get further independently, but their work could arrive at review larger, more complete-looking and harder to challenge. The organisation had become more capable in ways that made shared context, testing and collective judgement increasingly important.
Successful AI adoption had not simply removed a limit. It had changed the balance of the system.
This is the central mechanism of constraint migration: solving one problem changes which other problems matter.
Constraint migration becomes especially interesting when the new bottleneck sits well outside the technology.
In earlier work with a business managing a substantial portfolio of renewable energy assets, the ambition was to improve the monitoring and optimisation of wind and solar infrastructure. By combining operational data with predictive models, the organisation could develop a better understanding of asset performance, anticipate potential failures and identify interventions that might improve output or reduce downtime.
The technological capability created real opportunities. But some of those opportunities encountered constraints that better prediction alone could not resolve.
Consider a representative example from that work. A model predicts that a component is likely to fail and identifies a preventative replacement that could avoid a more expensive outage during a period of high generation. From the perspective of total asset performance, the recommended intervention makes sense.
However, the commercial arrangements governing that component may distribute costs and benefits between different parties. The manufacturer might have to pay for a replacement, while the asset owner captures most of the financial benefit from avoiding the outage.
The model has identified an economically valuable action, but the people and organisations responsible for implementing it may have no corresponding incentive to do so.
At this point, improving the prediction model may deliver little additional value. The system already knows enough to recommend the intervention. What now limits progress is the commercial arrangement through which that recommendation must become action.
Capturing the opportunity may require different incentives, contractual arrangements or participation from stakeholders who were not central to the original technology programme.
This is constraint migration from technical capability to commercial design.
The misalignment was not necessarily caused by AI. It may have existed for years, embedded in contracts that were designed around the information and capabilities available at the time.
As the technology improved, that arrangement became more consequential. The organisation could now see opportunities that it had previously struggled to identify, and the limits on acting upon them became harder to ignore.
The next model didn't need to be more intelligent. The commercial relationship did.
The software and renewable-energy examples illustrate different forms of the same underlying mechanism. But constraint migration can appear across many parts of organisational work.
One common pattern is a shift from production to judgement. As AI makes it easier to generate analysis, recommendations and strategic options, the organisation must decide which outputs deserve attention, which can be trusted and which are worth acting upon. Producing alternatives becomes less expensive, while evaluating their relative merits may become more demanding.
A second pattern is a shift from individual capability to coordination. AI allows individuals to undertake work that previously required a wider team, reducing dependence on specialist resources and accelerating execution. But greater independence can also weaken the points at which assumptions are challenged, context is exchanged and separate activities are reconciled into a coherent whole.
Another shift occurs from execution to verification and learning. When AI-assisted teams can produce and implement changes more quickly, the capacity to test those changes, identify unintended consequences and preserve lessons becomes more important. If the organisation continues to measure success primarily through the volume or speed of production, its ability to learn may fail to keep pace.
Constraints can also migrate from technical performance to authority and incentives, as in the renewable-energy example. The system can identify a better course of action, but existing decision rights, funding arrangements or commercial relationships may prevent the organisation from pursuing it.
These are not fixed stages through which every AI programme must pass. Nor should they be understood as an exhaustive taxonomy of organisational bottlenecks. The same intervention may alter several constraints simultaneously, and an improvement in one place can reduce or increase demands elsewhere.
The purpose of identifying these patterns is to develop a more useful diagnostic habit: whenever a capability improves, examine what has become limiting next.
Constraint migration is closely connected to another concept we explore in The Last Mile of AI: outcome throughput.
Task productivity describes how efficiently a bounded activity is completed. Outcome throughput concerns the rate at which the wider system converts connected activities into results that matter.
The distinction becomes important when AI produces substantial local productivity gains that do not translate into comparable improvements in organisational performance.
In one organisation we encountered, research and analysis that previously took around eight days could be completed in less than two with AI assistance. The productivity improvement was substantial, but the decision supported by that work still took roughly four weeks.
The original activity had become dramatically faster. The wider outcome had barely moved.
The work still needed to be interpreted, evaluated against competing priorities and translated into a decision. Those activities had become more important to the overall elapsed time because the analysis itself was no longer consuming as much of it.
In another case, AI-enabled work increased output in one part of an organisation by roughly three times, while the improvement in outcome-aligned throughput was estimated at around thirty per cent. Additional production created greater demands for verification, reconciliation, coordination and judgement.
Neither result meant that AI had failed to improve productivity. It meant the organisation's capacity to produce work had grown faster than its capacity to convert that work into outcomes.
Constraint migration helps explain why the gains can diverge. Once one part of the system becomes faster, another may govern how much progress is possible.
The practical implication is that measuring productivity only where AI is deployed can miss what happens to the work afterwards. A saving recorded in one team's workflow may create additional demands elsewhere, or it may produce genuine capacity that remains unused because another condition limits the result.
To understand the value of an AI intervention, we need to follow the work beyond the activity that became faster.
There is also an economic consequence to this shift.
Much of the visible discussion about AI economics focuses on production costs: the price of models, infrastructure, licences, tokens, compute and the human effort required to produce an output.
These are important measures, but they do not fully describe the economics of intelligence inside an organisation.
In The Last Mile of AI, we explore this through a framework we call Intelonomics, which considers the costs of producing intelligence, absorbing it and dealing with the consequences of what happens next.
As production becomes cheaper, the costs of reviewing, interpreting and acting upon what is produced may become relatively more important.
Imagine an organisation that can now generate credible strategic proposals in days rather than weeks. Teams begin producing substantially more options, but every additional proposal still requires leadership attention, evaluation, investment decisions and potentially a claim on limited execution capacity.
The organisation may have reduced the cost of creating each proposal while increasing the total demands placed on its decision-making system.
Of course, more options are not inherently a problem. They can improve strategic choice, expose opportunities and reduce the risk of overlooking a better alternative. But those benefits depend partly on whether the organisation can evaluate and act upon the additional intelligence effectively.
When production becomes abundant, value often depends increasingly on the capabilities that remain scarce.
Constraint migration is therefore not just an operational concern. It helps explain why the economics of AI transformation cannot be understood solely through model costs, hours saved or individual task performance.
The relevant economic transaction is larger than the AI interaction. It includes what the resulting intelligence requires from the organisation and what happens because the organisation acts, delays or chooses not to act.
The rise of agentic AI makes these dynamics more consequential.
Traditional AI applications often help people produce information or recommendations. Increasingly, agents can also plan activities, interact with tools, coordinate workflows and execute actions with varying degrees of autonomy.
This may remove genuine organisational friction. A process that once required several hand-offs can become substantially faster or, in some cases, largely automated.
But removing a hand-off does not necessarily remove the assumptions and authority structures it carried.
An agent can execute an objective that was poorly framed, optimise a measure that does not adequately represent the desired outcome or follow an outdated rule with extraordinary efficiency. It can also act faster than the organisation's existing mechanisms for review, correction and learning were designed to accommodate.
As more decisions and actions become automated, the effective constraint may shift towards how objectives are established, where authority is granted, how exceptions are handled and how behaviour is corrected when circumstances change.
Human judgement does not automatically become less important simply because fewer individual actions require direct human involvement. Some forms of judgement may become more consequential, particularly those concerning what the system should pursue, what risks are acceptable and when its behaviour requires intervention.
Agentic AI may therefore remove important parts of the work traditionally associated with turning intelligence into action, while exposing new limits in the organisation's capacity to direct and govern that action.
The Last Mile does not necessarily disappear as agents become more capable. Its constraints change.
The practical response begins with a change in how organisations approach transformation.
Instead of treating the original AI use case as a fixed problem to be solved and closed, leaders need to examine how the system changes once an intervention has succeeded.
The first step is to establish the outcome that matters. What is the organisation actually trying to improve, and how is that outcome produced today? This requires following the work across activities, teams, decisions and dependencies rather than limiting the analysis to the task that AI will perform.
Next, identify the condition currently restricting progress. That may be an expensive analytical activity, limited technical capacity, inadequate information or a process that genuinely needs to be accelerated. The diagnosis should be grounded in how work actually happens, not simply how the process is formally described.
The organisation can then make an intervention and evaluate its effect. If AI improves the targeted capability, the next question is not only whether the task became faster or cheaper, but whether the intended outcome changed.
Where improvement is smaller than expected, investigate what now limits the system. The problem might remain technical, or it may have moved into judgement, coordination, testing, incentives, authority or another organisational capacity.
This requires a willingness to reconsider the original intervention. Once the constraint has moved, continuing to invest in the capability that was previously limiting may offer diminishing returns.
A successful transformation programme therefore needs mechanisms for repeated diagnosis and adjustment, rather than assuming that its initial roadmap will remain the right answer as the organisation changes.
This does not imply endless restructuring or a perpetual programme of organisational change. Some interventions will create durable improvements without requiring substantial redesign elsewhere. The point is to recognise when the nature of the problem has changed and respond accordingly.
Constraint migration exposes a limitation in the way organisational transformation is often managed.
Programmes typically begin with an assessment of the current problem, followed by an agreed solution, funding, implementation milestones and measures of completion. That approach is entirely reasonable when the problem remains relatively stable and the intervention has predictable consequences.
But AI can change the capabilities of the organisation quickly enough that the original diagnosis becomes less useful before the programme is complete.
A team that began with insufficient analytical capacity may later struggle with prioritisation. An organisation that focused on software delivery speed may discover that testing and integration now control progress. A business that invested in better prediction may find that its commercial arrangements prevent it from acting on what it can now see.
In each case, the organisation has changed because the technology worked.
This suggests a different approach to transformation, one built around repeated cycles of understanding, intervention and learning. Rather than assuming that all the necessary changes can be specified in advance, organisations develop the capacity to identify emerging constraints and adapt their arrangements as new capabilities become available.
That capacity becomes increasingly important as AI itself continues to change. The tools, economics and feasible boundaries of automation will not remain static. Organisational arrangements that work well with today's capabilities may need to evolve as those capabilities improve.
The objective is not to redesign everything continuously. It is to develop an organisation capable of recognising when its existing design has become the constraint.
AI is making many forms of cognitive work faster, cheaper and more widely accessible. That creates substantial opportunities to improve productivity, accelerate decisions, expand organisational capabilities and pursue outcomes that were previously difficult or uneconomic.
But improvements to individual capabilities do not automatically produce equivalent improvements in the systems around them.
Sometimes AI removes a constraint and the whole organisation benefits. Sometimes it reveals another bottleneck, or shifts the limiting condition into an area that was not part of the original technology programme.
The distinction matters because the next useful intervention may have very little to do with making the AI more capable.
It may involve better judgement, clearer authority, stronger coordination, different incentives, improved commercial arrangements or a greater capacity to learn from what happens.
Constraint migration describes what happens when improving a capability changes which organisational weakness matters most.
Understanding that mechanism helps explain why successful AI initiatives can produce disappointing business outcomes, why the next problem often appears outside the technology, and why organisational transformation cannot always be managed as a fixed sequence of implementation activities.
The more capable AI becomes, the more important it is to keep looking beyond the capability itself.
That's part of what we mean by the Last Mile of AI: not simply getting intelligence into the organisation, but understanding and redesigning the conditions through which it becomes consequential.
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.

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.
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.
