Navigating the Fog: Why the AI Productivity Paradox Calls for a New Policy Playbook

John H Howard, November 18, 2025

This Insight is drawn from notes prepared for the CEDA-UTS AI Productivity Event, 4 November 2025

There is a critical and growing gap between the perceived revolutionary capability of artificial intelligence and its measurable real-world outcomes. We are told that AI is transformative. But when we examine macroeconomic figures such as GDP per hour worked, the impact is difficult to discern. This apparent disconnect, famously dubbed the "Productivity Paradox," is a sign of deep, structural uncertainty, a "fog" that confuses firm-level strategy and risks misdirecting public policy.

This uncertainty, however, is not a symptom of technological failure. It is the defining characteristic of genuine, disruptive innovation. Because innovation is about fundamental change, its ultimate outcomes cannot be predicted. They can only be understood in retrospect. This presents an immense difficulty. If we cannot predict the destination, how can we possibly chart a course?

This is not a new problem. As eminent Cambridge economist Dame Diane Cole observed recently, "we are in a position similar to that of the Victorians, who learned more about the Industrial Revolution from novelists than from official statistics" [1].

We are staring at an old map, wondering why it does not show a new continent. The 'paradox' is not a lack of impact, but a "measurement lag" combined with a failure of our existing models to reveal relevant and timely data.

This creates a "dangerous situation" for policymakers, who must make crucial decisions about regulation, investment, and workforce development now, without adequate empirical foundations.

The Measurement Mirage

The fog of uncertainty begins at the macroeconomic level. Our traditional economic metrics are not equipped for the AI transformation. As Coyle argues, our national statistics were designed for an industrial age, when tangible goods dominated. They already struggle to capture the value of the digital economy, let alone the impact of AI.

GDP, a lagging indicator, is poorly equipped to measure improvements in quality, variety, and customisation. An AI that helps a doctor improve diagnostic accuracy creates immense value that is not fully reflected in the cost of the hospital visit.

Many AI-driven gains are delivered as free services. One economist noted that trillions of photos are now produced at zero cost, a massive increase in consumer value that is invisible to GDP, which measures market transactions.

This problem has strong historical precedents. General-purpose technologies like steam power and electricity took decades to appear in national productivity statistics. This was not because they failed, but because firms first had to make costly investments in intangible assets: new business processes, employee retraining, and complete organisational redesign. The digital boom of the 1990s showed a similar lag, with productivity gains only appearing after years of sustained investment and integration.

This measurement gap creates a "data vacuum". We are "measuring blindfolded". We lack the most basic, systematic data on which firms are adopting AI, in which sectors, and for what specific purposes and applications. This vacuum undermines productive public debate, which fixates on fear rather than evidence.

Confusion on the Ground: The Firm-Level Myths

This macro-level fog cascades directly into firms, creating confusion and allowing persistent myths to take hold. Because leaders are also flying blind, they revert to simple, flawed, and often counter-productive metrics to justify their AI investments.

Perhaps the most common error is confusing productivity with efficiency. Many firms mistake efficiency, simple cost-cutting or speed for true productivity. This focus on cost-reduction misses the larger, transformative potential of AI to create new value, higher quality, and innovation, which are much harder to measure.

This leads directly to the myth that productivity equals headcount reduction. This narrow view equates AI's value purely with automation. It completely overlooks the more significant and often more valuable impact of augmentation. Real gains are frequently found when AI enhances the skills of the existing workforce, allowing them to solve more complex problems or create new services.

An example illustrates this perfectly: if a law firm measures AI success by a 15% reduction in paralegal staff, and meets this target, it is likely that senior lawyers will spend extra hours correcting the AI's subtle contractual errors. The firm probably missed the real productivity gain: using the AI to augment its paralegals, reassigning them to higher-value client research that would have increased billable capacity and client satisfaction.

This reveals another myth: treating AI as a simple 'plug-and-play' IT solution. Many organisations delegate AI to the IT department like a software update. This fails to recognise that leveraging AI is a strategic business transformation. Real productivity gains require fundamental, top-led changes to workflows, business models, and employee skills.

We may have situations where a manufacturing company buys an AI for predictive maintenance, and the IT group installs it, but does not take the time to consult line managers and crews. The crews, lacking training, find the AI's alerts disruptive and do not trust them. They revert to their old manual schedules, and the expensive system sits unused. The firm treated AI as an installation, not a change to workflow and culture.

Even when firms measure "time saved," they often ignore the creation of new, inefficient work, dubbed "workslop". This includes managing new complex workflows, duplicated tasks, and the manual correction of flawed AI outputs. This new, unmeasured work can negate the time saved. This is particularly true for high-skilled professionals, who must spend substantial time correcting subtle AI errors, offsetting speed benefits.

Two final myths compound these errors. First is the belief that more data automatically means better AI. This ignores the "garbage in, garbage out" principle. Using poor, biased, or irrelevant data creates flawed outputs that result in negative productivity, as employees must correct the system's mistakes. A retailer using 20 years of messy sales data to forecast new fashion will get deeply flawed, costly recommendations.

Second is the assumption that gains are universal and transferable. Productivity is highly context-dependent. An AI chatbot that succeeds at a high-volume bank will fail miserably at a high-net-worth financial advisory firm, whose clients value personal, bespoke advice. The tool's value depends entirely on the specific business context, tasks, and skills.

A New Playbook for Navigating the Fog

The "most dangerous misconception" is that we can afford to wait for better data before acting. This is not an option. If outcomes are unknowable, our entire policy and strategic framework must be rebuilt around navigation and adaptation. A new Playbook would address the following.

1. Reframe the "Outcome" as a "Mission"

Much of our innovation policy thinking follows a "linear and physical process", a flow from "Mind to Market". AI and quantum technologies disrupt this entire linear model. AI allows for rapid design and testing in silico. The most valuable 'product' may not be a physical device but an algorithm, and 'scale' becomes less about a factory line and more about data.

Our established policy levers, like fabrication infrastructure, may no longer be the primary ones. Instead of defining a static solution, policy must define a dynamic mission. A mission provides a clear, ambitious goal but grants the flexibility for innovators to discover the non-linear, unpredictable pathway to achieve it.

2. Measure the Transformation, Not Just the Result

If we cannot know the final outcome, we must measure the process of transformation itself. Instead of relying on lagging indicators like GDP, we must invest in leading indicators. This is the practical answer to the measurement problem. We must fund our statistical agencies and academics to build the "fundamentally new categories" of measurement that are required.

These new metrics would track the change as it happens: systematic data on AI-related energy consumption, specific labour market shifts, time-use data in workplaces and homes, and structural indicators like changes in industrial composition and organisational design. This is how we begin to penetrate the fog.

3. Build the System, Not Just the Tool

If we cannot predict the winning product or firm, we must invest in the system that produces and supports them. This is the core hypothesis of the UTS project for Google.org: the quality of an innovation ecosystem is a primary factor explaining the AI productivity lag. The most important "outcome" for policymakers to manage is the health, resilience, and adaptive capacity of this system.

AI adoption accelerates, and productivity impacts become observable in places that successfully align placemaking, economics, business, and governance with shared infrastructure. The policy goal must be to assess ecosystem readiness, diagnose gaps, and make targeted interventions that allow AI pilots to effectively scale. This builds the national capacity to absorb the technology.

4. Invest in the "Critical Lens"

Finally, we must invest in the human skills required to navigate this new landscape. As AI evolves from a "co-pilot" to more autonomous "agentic" systems, the value of human professionals will shift. It will move from doing the work to designing and managing new, augmented work systems.

Beyond basic data literacy, the most critical skills will be in evaluation, governance, and systems-level thinking. We must train professionals to be the "critical lens". Their value will be the ability to interpret, question, and govern these new processes, ensuring they align with organisational and societal goals.

Conclusion: Prescience, Prediction and Preparedness

The AI Productivity Paradox is not a failure of technology. It is a profound failure of our existing models of measurement, strategy, and policy. It is a clear signal that we are in the early, messy, and uncertain stages of a true general-purpose technology diffusion, just as our predecessors were with steam and electricity.

The challenge for leaders is not to find a crystal ball. The challenge is to build a better compass, a more resilient ship, and a more capable crew. The task is to stop trying to predict the future and start building the national and organisational adaptive capacity to navigate it as it unfolds.

Reference.

Diane Coyle, Professor of Public Policy at the University of Cambridge, is the author, most recently, of The Measure of Progress: Counting What Really Matters (Princeton University Press, 2025).

Note.

[1] “Measuring AI’s Economic Impact”, Project Syndicate, 21 October, 2025