Dharmateja Priyadarshi Uddandarao
The real story of the AI economy is not the data. It is the judgment required to decide what the data actually measures, and that judgment is the statistician’s craft.
Economist Rebecca Riley writes in the December issue of Finance & Development that assessments of the world’s economies may be off by trillions of dollars. “GDP, CPI, and total factor productivity for aggregate economy-wide are struggling to keep up with the speed of change of technology, business models and consumer behavior in a data-driven economy.”
Measuring the economy to determine whether to encourage growth or control inflation to keep prices from rising requires good economic data. Understanding where different communities are in terms of economic well-being and which industries are producing value in their economies and where mere numbers are being shifted from one location to another requires good economic data.
Additionally, in many economies, now more than half of intangible investments are data-related and only recently started to be included in the economic accounts. Large portions of companies’ investments in software, in algorithms, in organizational capital, in AI-related systems and activities are not captured by the accounting systems designed to track the physical stock of factories and physical goods.
The gap between what data shows and what data means for users is a fundamental challenge in working with data. For none of the areas where data is used do the tensions surrounding working with data seem to matter more than they do for economic statistics, which inform decision-making by central banks, fiscal authorities, and investors influencing billions of people worldwide.
This is where the work of a statistician comes in. Economic statistics are not facts that can be collected. Rather, they are estimates based on several decisions. The choices of what to measure and how to measure it all must be taken, and an assessment of the uncertainty involved in the estimates must be made. As the economy evolves at a pace at which the choices made regarding what to measure and how cannot keep up, it will be necessary to ask whether what one is measuring is what one thinks one is measuring. Questions such as these are what a statistician brings to a study of the AI economy.
Four Statistical Problems Hiding in Plain Sight
Four major problems need to be addressed behind the headline “measurement gap,” and they all need a good statistician to solve them.
First, unobserved activity. National accounts do not have a line item for the value of AI. So, we must observe its impact indirectly. And, as with all latent variables, the statistics about them are necessarily rooted in modeling assumptions and thus subject to wide latitude, just as the epidemiologist’s best estimate of an unseen exposure is based on the trace it leaves.
Second, prices and quality may move together and thus be observed only in combination. For instance, a laptop that has the same price but is twice as fast as before may indicate that prices for laptops have fallen or that the quality of laptops has increased. A statistician would make a model to separate these two effects, and different reasonable choices about which characteristics to include would result in significantly different measured growth.
Third, zero-price goods. This is the familiar problem that a good with zero price does not equal a zero value. Many free online apps and services are saving millions of hours per week, and much of this is not being recorded in GDP. These are not zero value activities, and statistical models are needed to value them, such as willingness-to-pay experiments or choice models, to record value on accounts that only record transactions on invoices.
Fourth, identification under structural change: AI-induced changes are structural in nature; thus, we cannot use past data for predicting future ones. Adopting firms are not a random sample; the ground under their feet is shifting as other forces affect them simultaneously. The job of a statistician here is to establish whether the AI-induced effect can be identified at all.
And, through all of these, judgment (i.e., whether one can trust a number or not) plays a major role.
AI Makes the Problem Worse Before It Makes It Better
One expects an economy increasingly running on data, processing it ever faster, with more information at one’s fingertips all the time, must increasingly be able to be quantified and measured. And in fact, AI does promise all this.
Table 1—Economic Measurement: Traditional Frameworks vs. AI-Era Realities
| Dimension | Traditional Framework | AI-Era Reality |
|---|---|---|
| Primary assets measured | Factories, equipment, inventory | Software, algorithms, organizational capital |
| Investment visibility | Capital expenditure on balance sheets | Intangible spending often expensed, invisible in GDP |
| Price measurement | Physical goods with observable quality | Digital services consumed for free or bundled |
| Productivity signal | Output per hour, directly attributable | Indirect, delayed, distributed across functions |
| Data latency | Quarterly or annual surveys | Real-time digital traces available but unintegrated |
| Causal attribution | Before-and-after policy comparison | Entangled interventions requiring counterfactual methods |
The Bureau of Economic Analysis is doing important work to figure out how to measure the influence of AI and the contribution of AI to US macroeconomic accounts. For now, there is no direct line item for AI in the US national accounts, and their findings must be inferred from indirect evidence. For example, their baseline model of labor market activity found AI was associated with higher productivity and a reduction in the use of inputs, including labor, but these results were considered preliminary and subject to a wide range of assumptions. IT investments increased in cost before generating increases in revenue. And even though everyone knew IT was transformational in terms of operations, there were years of further research by economists required to track its contribution to GDP.
Just as earlier technologies such as the IT of the 1990s had to wait for years for their economic impact to be quantified by productivity statistics that had been designed for a completely different purpose, AI too will have to wait before its impact can be accurately recorded by economic statistics. Not least because of a special feature of this new technology: Unlike IT of the 1990s whose economic effects had been obvious to firms long before they could be measured by statistics, today’s investments in AI first increase costs before any revenues will be generated—indirectly and in unforeseeable quantities. And it is precisely this special feature of AI that also makes the measurement of its economic effects so particularly complex.
From Prediction to Intervention
In economics, predicting GDP growth of 2.3% in the last quarter is not enough. We want to know how this came to pass, how some other policy would have worked, and how long this growth is going to last. These are all causal questions and require the kind of judgment a good statistician can bring to bear on the data.
A real-world problem for policymakers is that the regions with the highest AI adoption also have the highest share of younger, less-educated workers. This could be the case of AI pushing out experienced workers or AI opening doors for new workers to enter. In both cases, the chart appears to be convincingly supportive of a particular policy direction, but correlation is not causation. To make good policy decisions, the correlations must be unpacked and the lines of causation determined. This is what good statisticians do for a living.
Specific methodologies help answer specific questions. Quasi-experimental designs test for the counterfactual change that would have occurred in the absence of the policy change under consideration. They are particularly faulty when the groups being compared are never identically situated. Quality-adjusted price indexes try to strip away changes in price that reflect changes in quality rather than cost. However, they are prone to error if the set of characteristics used to measure quality is not optimal. Alternative data to official national accounts contains a host of representative errors (e.g., nonrepresentativeness of the sample). Measures of uncertainty for any given statistic can highlight the extent to which that number is likely to be wrong and thus serve as a useful check on false precision. And counterfactual models of what would have occurred in the absence of a policy intervention are essential for making sensible policy decisions.
What any method can do wrong is far more valuable to understanding the economy than what it can do right. The use of a particular tool to ask a question is not as important as the questions that are asked using it.
Why the Statistician Is Central
A popular conundrum being watched by many within the profession is outlined in a July 2026 FEDS Note. That note organized the ‘many indicators of the influence of AI’ into three groups: the capabilities and costs that firms are investing in; the investment and adoption that follows; and the changes in output and labor resulting from that investment and adoption. Notably, the note identified a wide gap between the responsiveness of financial markets to the AI narrative and the data on aggregate output and labor.
Instead of rushing to answer the question of a puzzle the profession is currently following (The influence of AI – a FEDS Note from July 2026 Organizing public indicators into three areas: 1. the AI capabilities and costs that firms are investing in; 2. investment and adoption by firms; and 3. changes in productivity and labor resulting from that investment and adoption), the statistician sits with the puzzle for a while. The puzzle has three possible answers: the transformation has not yet taken place; there are other forces at work counter-balancing its impact; or the statistics we have do not capture what is already happening.
This is where the statistician’s way of thinking becomes interesting to follow. First, is the data really measuring what one thinks it is? A statistic, labeled as “AI investment” or “AI productivity” is only as good as the definition behind it. Then, inspecting the assumptions of a model, looking at parallel trends, checking for stable relationships, and checking for a representative sample as opposed to blindly “churning out” numbers and inserting them into a pre-built model becomes important.
Then, there is the issue of correlation vs. causation, which is probably the biggest pitfall of them all when working with data. After that, it’s a case of honest uncertainty assessment and providing a range with caveats as opposed to blindly churning out numbers that can end up being wrong.
Skeptical scrutiny, of course, is not something that can be “automated” or even enhanced using better software. With more data and more computing power available, the role of a statistician is not to pass along smooth and confident forecasts to decision-makers, but rather to warn them about the use of numbers—by machines or humans—that deserve not to be trusted.
Statisticians are the mapmakers of the modern economy. When the territory shifts, they are the ones who redraw the map so the rest of us do not get lost.
Statisticians are the mapmakers of the modern economy. When the territory shifts, they are the ones who redraw the map so the rest of us do not get lost.
Why This Matters Now
A nation’s economy is growing or contracting by trillions of dollars, and much of this activity often goes unmeasured or is measured inadequately. While the economy is rapidly transforming due to the influence of AI, our statistical systems are struggling to keep up. Decisions will be made based on these measurements for years to come, influencing interest rates, welfare, business contracts, and the national budget.
Data scientists and statisticians recognize that the same causal reasoning skills that can determine the optimal launch strategy for a new product or drive higher returns on investment by optimizing processes also need to be brought to bear on economic measurement. Such analysts would bring a healthy dose of skepticism to bear on the headline numbers reported by national statistical offices, questioning whether the indicators reported upon (whether it be unemployment, inflation, GDP, etc.) are in fact being measured in a manner that is comparable across space and time.
We must ask the same questions of the economic statistics that measure the economy of an entire society that we ask of any other dashboard: Are we measuring what matters, or just what’s easy to count?
Leave a Reply