What is Outcome Throughput?

Why AI productivity doesn't always translate into organisational progress.

By Jordan Dalladay and Craig Wright | UN/DO

One team we encountered had achieved the sort of AI productivity gain most organisations are hoping for. A piece of research and analysis that previously took around eight days could now be completed in less than two. The work arrived earlier, the team had more capacity, and a substantial amount of effort had been removed from the process.

There was no reason to diminish the achievement. The technology was doing exactly what it was supposed to do, making a valuable activity dramatically more efficient.

But when we followed the work further, something interesting emerged. The decision that the analysis supported still took around four weeks. The research had become faster, but the outcome had barely moved.

The AI had delivered a genuine productivity improvement. What it hadn't necessarily delivered was a corresponding improvement in organisational performance.

That distinction is what we mean by outcome throughput, one of the central ideas explored in our book, one of the central ideas explored in our book, The Last Mile of AI.

What is outcome throughput?

Outcome throughput is the rate at which an organisation converts connected activities into meaningful results. Unlike task productivity, which measures the efficiency of an individual activity, outcome throughput considers the wider system of work, judgement, decisions, coordination and action through which value is actually created.

The difference matters because most organisational outcomes are not produced by a single task. They emerge from interdependent activities involving different people, functions, systems and decisions, often operating on very different timescales.

A research report may inform a strategic recommendation, which needs to be evaluated alongside competing priorities before becoming an investment decision. That decision then needs to be translated into operational changes, coordinated across teams and ultimately reflected in customer, commercial or organisational outcomes.

Making the report faster can help, sometimes enormously. But the value of that improvement depends on what happens throughout the rest of the system.

In The Last Mile of AI, we distinguish between two forms of productivity:

Task productivity is the efficiency with which a bounded activity is completed.

Outcome throughput is the rate at which the wider system converts connected pieces of work into a result that matters.

The distinction is not an argument against productivity. It's an argument for understanding productivity at the level where its consequences become meaningful.

A faster task is not necessarily a faster system

Most organisations have good reasons for measuring productivity where they do. Tasks have relatively clear boundaries, making it possible to measure how long something takes, how much it costs or how many units of work are completed. Outcomes are harder to isolate because they emerge from multiple activities, and the eventual result may be affected by factors beyond the organisation's immediate control.

AI makes this measurement problem more consequential because its improvements to individual activities can be so dramatic. Research that took days becomes a matter of hours. Documentation that required extensive manual effort can be generated in minutes. Product teams can explore more options, engineers can produce more code and analysts can investigate questions that previously required specialist support.

These are meaningful capabilities. But they can also create a misleading picture of overall progress.

Consider a product organisation in which AI dramatically accelerates the creation of requirements documents. Previously, developing a detailed specification required several days of research, discussion and drafting. Now a product manager can produce something comprehensive and convincing in an afternoon.

The team can generate more specifications than ever before, but that doesn't mean it can develop, test, release and learn from more valuable products at the same rate. The engineering team still needs to interpret the requirements, competing priorities still need resolving and the organisation still has to decide which propositions deserve investment.

In fact, the apparent productivity gain may reveal a deeper problem. The effort involved in writing requirements had sometimes forced difficult conversations about priorities, assumptions and product direction. Generating the document more quickly can remove that friction without resolving the questions the process was helping people confront.

The visible task has become easier, but the work of reaching a shared understanding hasn't necessarily disappeared.

This is a familiar principle in systems thinking. Improving the performance of one component does not automatically improve the performance of the whole system, particularly when that component was not the limiting factor.

The difference is that in knowledge work, the constraints are often harder to see. There may be no physical queue accumulating on a factory floor. Instead, work accumulates as decisions awaiting approval, recommendations requiring judgement, projects awaiting coordination or ideas competing for limited attention.

The queue is cognitive.

When more output creates more work

In another organisation we encountered, AI-enabled working allowed one part of the business to produce roughly three times as much work as before. Looked at locally, the improvement was extraordinary. Yet the increase in outcome-aligned throughput was estimated at only around thirty per cent.

Again, the productivity gain was real. The organisation had genuinely increased its capacity to produce work, but its capacity to absorb that work had not expanded at the same rate.

More outputs needed verification. More material had to be reconciled with work happening elsewhere. More options required comparison, and more recommendations eventually needed decisions. The organisation had become much better at generating intelligence without becoming proportionately better at converting that intelligence into action.

This exposes an important characteristic of AI-enabled productivity: the cost of producing work and the cost of absorbing it can move independently.

When producing an analysis is expensive, the effort naturally limits how much analysis enters the system. When production becomes cheap, the organisation may receive more material than its existing mechanisms for judgement, coordination and decision-making were designed to handle.

Sometimes this is valuable. More options can improve decisions, better analysis can reveal previously invisible opportunities, and greater capacity can allow organisations to address problems they had neglected.

But abundance doesn't automatically create value. More output can also produce more reviewing, more meetings, more competing proposals and more opportunities for work to become fragmented.

The organisation may save time in one place while creating additional demands somewhere else. What disappears from one team's timesheet can reappear in another person's calendar.

This is why we should be careful about treating hours saved as a complete account of productivity. The more useful question is whether work was genuinely removed, whether it was transferred elsewhere or whether the additional capacity created something of value.

The difference between producing intelligence and using it

The distinction becomes even more interesting when the work being accelerated produces insights rather than finished products.

In one engagement, we worked with a commercial strategy team whose process for developing customer personas relied on extensive research and behavioural synthesis. Done properly, the work could take weeks or months. With AI-assisted research and synthesis, much of that effort could be compressed into days, creating the possibility of refreshing customer understanding far more frequently.

At first glance, this was a compelling productivity opportunity. Better information could be produced more quickly, potentially giving the organisation a much more responsive view of its customers.

But the team wasn't particularly effective at using personas at the old speed either.

Creating them faster did not automatically mean people could translate those insights into better propositions, evaluate opportunities, challenge assumptions or design more effective customer experiences. The organisation had improved its ability to produce intelligence without necessarily improving its ability to act on what it discovered.

That changed the nature of the intervention.

Rather than continuing to optimise persona generation, we developed a design sprint approach supported by an AI simulator that helped teams explore and apply the customer perspectives.

The opportunity moved from producing the insight to helping the organisation use it.

This is an important distinction because so much AI investment begins with identifying activities that can be accelerated. The obvious target is often the production of an artefact: a report, a presentation, a forecast, a specification or an analysis.

But the artefact isn't necessarily the outcome. Its value depends on what it helps the organisation understand, decide or do differently.

An organisation can be exceptionally productive at generating insights and remarkably ineffective at turning them into consequences.

The constraint moves

Once we start looking at outcome throughput, another pattern becomes visible: improving one part of the system can change which part of the system limits progress.

We call this constraint migration.

In the eight-days-to-two-days example, the time required for analysis fell dramatically, making the decision process more consequential to the overall outcome. In the three-times-output example, the capacity to produce work expanded faster than the capacity to review, reconcile and act on it. In the customer-persona example, generating insights became easier while activating those insights remained difficult.

In each case, improving one capability changed the relative importance of the surrounding constraints.

This is not necessarily evidence that the AI intervention failed. Sometimes it's evidence that it worked well enough to reveal a limitation the organisation could previously afford to ignore.

The mistake is to assume that the next investment should continue improving the same activity.

If research was once the bottleneck, it may be tempting to make research faster again. But once the organisation can produce analysis in two days rather than eight, the next meaningful improvement may lie in how decisions are made, how priorities are agreed or how responsibility for action is distributed.

The technology may have successfully solved the problem it was asked to solve. The organisation now has a different problem to address.

Understanding outcome throughput therefore requires more than measuring the initial productivity gain. It requires following the work far enough to discover what has become limiting next.

Why AI ROI can be difficult to see

This also helps explain why organisations can accumulate impressive AI productivity statistics without seeing an equivalent improvement in business performance.

The metrics commonly used to demonstrate AI value tend to be close to where the technology is deployed: hours saved, tasks automated, documents generated, adoption rates, model costs or the speed of individual workflows.

These measures are useful. They tell us whether a particular activity has become more efficient or whether a capability is being used. The problem begins when they are treated as evidence of value at a much wider level.

A forty per cent reduction in the time required to produce an analysis is not necessarily a forty per cent improvement in organisational productivity. Nor does it imply a corresponding reduction in operating costs, an increase in revenue or an improvement in customer experience.

The benefit depends on what happens to the capacity created and whether the improvement changes the wider system.

Sometimes the effect will be substantial. If an activity was genuinely constraining the outcome, removing that constraint may accelerate the whole process. In other cases, the benefit may be modest because the activity represented only a small part of the total elapsed time.

And sometimes faster production creates additional work elsewhere, offsetting part of the apparent saving.

In the book, we explore this through a wider economic lens we call Intelonomics, distinguishing between the cost of producing intelligence, the cost of absorbing it and the economic consequences of what happens next.

Outcome throughput helps make those distinctions operational. It shifts attention from the cost of generating an output towards what the organisation is actually able to achieve because that output exists.

That doesn't make measurement easier. But it makes the question more useful.

How should organisations measure outcome throughput?

There is no universal outcome-throughput metric that will be appropriate for every organisation. What counts as a meaningful result depends on the system being examined, and different outcomes have different timescales, quality requirements and constraints.

For a product organisation, the relevant outcome might be the rate at which valuable improvements reach customers and produce measurable changes in behaviour. For a commercial team, it might be how effectively customer insights are converted into validated propositions or growth opportunities. For an operational business, it may concern how quickly and reliably problems are identified, resolved and prevented from recurring.

The important thing is to make the relationship between activity and consequence explicit.

A useful approach begins by identifying an outcome that matters, then tracing the connected work required to produce it. Where does time accumulate? Where do decisions wait? Which activities generate value, and which primarily create work for other parts of the organisation? What happens to the outputs AI produces once they leave the immediate task?

This makes it possible to measure productivity at more than one level. Rather than abandoning local efficiency measures, we can examine them alongside broader indicators of system performance.

A research team might track how quickly an analysis is produced, while also examining the elapsed time from the original question to a consequential decision. A software team might measure coding speed alongside release throughput, quality and rework. A strategy team might examine the number of options developed, but also the rate at which worthwhile opportunities are validated, funded and acted upon.

There is an important caution here. Faster outcomes are not always better outcomes. Good judgement may require deliberation, testing or additional evidence, and not every decision should be accelerated.

The objective is not maximum organisational speed. It is improving the system's ability to produce outcomes that matter, at an appropriate level of quality, cost and risk.

Outcome throughput is therefore as much a diagnostic perspective as a performance measure. It helps leaders distinguish genuine progress from local acceleration and identify where further improvement would make the greatest difference.

Start with the outcome, then follow the work

Most AI transformation programmes begin with a familiar set of questions. Where could we use AI? Which activities are expensive or repetitive? What could we automate, and how much time might we save?

These are reasonable questions, but they start with the capability and search for places to apply it.

A Last Mile approach begins somewhere different. It asks what meaningful outcome the organisation is trying to improve, what currently constrains that outcome and whether AI has a useful role in changing those conditions.

The intervention may still be technical. Better models, data, interfaces or workflow integration may genuinely improve the result.

But the necessary change might instead concern decision rights, incentives, coordination, review capacity or the way people interpret and use information. Sometimes the most valuable intervention is not to generate more intelligence, but to improve the organisation's capacity to make sense of what it already has.

This is why we argue for starting with the Last Mile, rather than treating it as the final problem to solve after AI has been deployed.

Once an intervention has been made, the work continues. The organisation needs to examine what actually changed, whether the outcome improved and what new constraint has become important. That learning should inform the next intervention, rather than allowing the original use case to define the boundaries of the transformation indefinitely.

From productivity to progress

AI is changing the economics of producing many forms of cognitive work. Analysis, writing, coding, research and the generation of plausible options can now become dramatically faster and cheaper in the right circumstances.

That change creates extraordinary possibilities. But it also challenges an assumption that has shaped much of how we measure work: that producing more, or producing the same thing faster, is necessarily evidence that the organisation has become more productive.

Sometimes it is. Sometimes it isn't.

The difference becomes visible when we follow the work beyond the task itself, through the judgement, decisions, coordination and action required to create meaningful outcomes.

Task productivity tells us how much faster the work became. Outcome throughput tells us whether the organisation actually moved.

As intelligence becomes more abundant, that distinction will become increasingly important. The organisations that benefit most from AI may not simply be those that generate the greatest amount of intelligence, but those that become better at turning intelligence into consequence.

That's the challenge of the Last Mile.

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.