Why AI’s Investment Surge Has the Fed Watching More Than Productivity
AI investment is lifting demand before its productivity payoff is clear. The Fed’s November 13 assessment exposes a gap between spending and output.
Federal Reserve Vice Chair Philip Jefferson’s November 13 speech put the immediate demand from AI investment alongside its slower potential productivity gains. Speaking at UC San Diego, he identified data-center spending, equity-related wealth effects and pressure on memory and logic chips. This was his economic assessment, not a new interest-rate decision. Jefferson’s speech
Understanding that assessment requires separating three things that often share the same headline: buying capacity, using a tool, and producing more value. A server order establishes demand for equipment. A worker opening an assistant establishes adoption. Neither alone measures how many additional products or services customers ultimately receive. Confusing those stages makes both spectacular success claims and declarations of failure arrive too early.

An imported server can matter twice, in different ways
An equipment purchase raises a subtle question: where was the equipment produced, and where will it be used? The first question matters for current domestic output; the second matters for the capacity available to workers afterward. A large spending announcement does not answer both.
BEA explains that business investment includes purchases of imported equipment, while the imports entry removes foreign production from the calculation of US GDP. That adjustment prevents foreign-made goods from being counted as domestic output; it does not mean an imported machine destroys an equal amount of American production. BEA's expenditure-accounting explanation
Jefferson makes the related distinction in his speech's eighteenth footnote: imports satisfy a substantial part of IT investment demand, limiting the direct GDP effect, while imported capital can still raise labor productivity. Later, he describes possible upward pressure on the neutral interest rate if investment demand increases and saving falls, but stresses that the relationship is uncertain. He also distinguishes cyclical weakness from a persistent skills mismatch that monetary policy cannot repair. Speech and footnote 18
For readers following the AI buildout, the implication is concrete. Domestic equipment manufacturing, construction and subsequent operation are separate activities. A dollar total for the project is not a dollar estimate of its immediate contribution to US output. Yet looking only at that immediate contribution can miss the productive asset available once the project operates. Neither the investment headline nor the import subtraction is a complete verdict on the project.
The buildout also raises a separate policy question about who pays for electricity infrastructure. Our reporting on the Senate’s proposed data-center power deal examines that cost-allocation debate, rather than treating it as a measurement of AI productivity.

Count the work, not just the people who tried a tool
A closer look at research underlying the debate reveals how easily an adoption statistic loses its meaning. In What Work Does Generative AI Do?, Alexander Bick, Adam Blandin, David Deming and Tyler Schumacher use an online survey of US adults aged 18 to 64. The task module covers August and November 2025 and February and May 2026, rather than a newly conducted November survey.
For each respondent's occupation, the researchers identify leading tasks from O*NET, ask which the respondent regularly performs, and then ask about AI assistance. A task's usage denominator is the people doing that task. The survey uses demographic quotas and weights; that does not eliminate all selection concerns. This is a working paper, not a final national census of AI activity. Author manuscript and methods
Frequency introduces another distinction. Someone who tests a tool once and someone who uses it throughout each working day both qualify under a sufficiently broad question about use. A business surveying whether any department has adopted AI is asking something different again. Without the question, reference period and population, an adoption percentage is a label with its measurement instructions removed.
That changes what readers should demand from the next confident claim that AI is everywhere. Is the count of workers, employers, tasks, hours or software sessions? Does use mean experimentation, routine assistance or an autonomous process? Those are not academic footnotes. They determine whether a statistic describes curiosity, deployment or a substantial change in production.

The surprising date mismatch in an economic study
Research highlighted by the Bureau of Economic Analysis offers a particularly important limitation. Its AI-intensity measures cover the second quarter of 2025 through the first quarter of 2026, while the economic outcomes examined cover 2016 through 2024. Places and industries with higher measured AI use were associated with stronger earlier output performance. That is not a demonstration that their later measured adoption caused their earlier growth.
The study compares worker and employer measures and finds that results depend on the measure and level of aggregation. BEA's October 2 spotlight presents the work as descriptive. The underlying paper was published in August. Neither publication is a fresh November output release. BEA's explanation, underlying research
The chronology does not make the research useless. It changes the question it can answer. A later measure may help identify the kinds of industries or locations where adoption is concentrated. It cannot, by itself, establish when the productive change occurred. A business environment favorable to both growth and technology use could produce an association without the technology being the sole explanation.

Strong workplace results still have boundaries
There is evidence of meaningful improvement in a specific workplace setting. Research by Erik Brynjolfsson, Danielle Li and Lindsey Raymond examined the staggered introduction of an AI assistant among 5,172 customer-support agents. The reported average increase in issues resolved per hour was 15 percent. Results differed by experience and skill, with stronger benefits for less experienced workers. Stanford's research summary
That outcome is more informative than counting subscriptions because it measures completed work against time. But customer-support issues are not a universal unit of economic output. An increase in that measure cannot simply be applied to every occupation, multiplied by the entire workforce and presented as a national forecast. The result belongs to the studied system, workers and tasks.
An editor, a machinist and a customer-support agent can all use an assistant while facing different constraints. A draft may be produced quickly but still require checking. A manufacturing recommendation may depend on equipment availability. A support response may succeed only if the underlying customer problem is actually resolved. These examples illustrate different production chains, not measured estimates of their AI gains.

Output has to survive the rest of the workflow
The Bureau of Labor Statistics defines labor productivity through the relationship between output and hours worked. More equipment, worker skills, organization and other changes can contribute. It is not a statistic that identifies the contribution of one technology automatically. BLS productivity overview
This distinction matters when a company reports sales per employee. Revenue can change because of prices, the mix of products sold or demand, as well as productive efficiency. A firm can also improve a service in ways that are difficult to capture by counting items. These measures can illuminate different parts of its performance, but substituting one for another would erase the question each was designed to answer.
The customer-support finding is useful precisely because it names an operational result: issues resolved per hour. Before applying that result elsewhere, an analyst needs to know what counts as completion in the next setting. For a coding assistant, producing lines is not the same endpoint as working software. For a document workflow, generating text is not the same endpoint as an accurate final document. The unit of completed work comes first; only then can a speed comparison mean much.

The smaller business's missing line items
For a smaller employer, the relevant cost is not just the monthly subscription. A hypothetical evaluation might compare 100 monthly tasks taking an hour each with a new process taking 40 minutes plus 10 minutes of checking per task. That reduces total task time from 100 hours to about 83 hours and 20 minutes. If operating and maintaining the process takes another 10 hours each month, the remaining saving is only about six hours and 40 minutes.
Now add 30 hours of one-time setup. Under those exact assumptions, the time saving would take about four and a half months to recover the initial effort. This is an illustrative calculation, not a claim about typical returns, labor prices or any actual business. It excludes subscription charges, error losses and the value of alternative uses of staff time. Changing the review requirement changes the result immediately.

The example shows why seemingly minor implementation details matter more than a dramatic headline percentage. A tool that saves time on a frequent, standardized task can have a different payoff from one that accelerates an occasional assignment. A process requiring expensive specialist review can produce a different result from one where a quick check is enough. A firm must know which situation it has before attaching a dollar value to the apparent saving.
There is a factual reason not to erase review from that calculation. NIST's generative-AI risk profile identifies confidently incorrect output and human overreliance among the risks organizations should evaluate. It recommends checking sources and citations as part of ongoing measurement. The profile is voluntary guidance, not a newly imposed November rule. NIST's generative-AI profile
A reliable workflow can still be highly worthwhile. But reliability is part of what the business is producing, not an optional surcharge to omit from its return calculation. The commercially useful comparison is a complete old process against a complete new one, with the same standard of acceptable work. That approach gives both a genuinely successful tool and a disappointing one a fairer test.

Why the household payoff will not arrive in one form
An economy can become more productive while households experience the benefits differently. Federal Reserve Governor Michael Barr's November 12 discussion of AI and inequality emphasizes how ownership, work and competition can affect the distribution of gains. Those channels complicate any assumption that a technological improvement reaches everyone through the same paycheck. Barr's speech
Consider three possible routes without assuming any is guaranteed. A business might share additional value through wages, retain it as profit, or pass some of it to customers through lower prices or better service. A worker's experience depends on which route operates and whether the person's tasks remain in demand. A consumer's experience depends on the products purchased. The same productivity improvement can therefore matter differently across a household's working life and shopping basket.
Borrowing adds a separate contractual layer. An existing fixed-rate mortgage does not reset its interest rate when a policymaker gives a speech. An adjustable-rate mortgage follows its contractual index, margin and adjustment provisions. Property taxes and insurance can still change a homeowner's total payment even when the mortgage interest rate is fixed. CFPB's mortgage explanation
The practical distinction is between today's contract and tomorrow's borrowing decision. A claim that AI will eventually reduce costs is not a promised refinancing opportunity. Nor is concern about demand from the AI buildout a notice that every household's payment will rise. Those outcomes require additional evidence about policy, markets and the specific contract. Turning a long-run technology forecast into an immediate household payment forecast skips those steps.

What would make the next update genuinely informative?
For the spending side, an announcement should be distinguished from equipment delivered and capacity put to use. For the output side, a faster task should be distinguished from additional completed work. For the household side, company profit, wages and consumer prices should be examined separately. These distinctions create a way to evaluate the next announcement without requiring either faith in a technological revolution or disbelief in any benefit.
The policy calendar also matters. The Federal Reserve lists November 18 for release of minutes from the October meeting and December 8–9 for its next scheduled policy meeting. A speech is one official's analysis; minutes describe an earlier discussion; the next decision will be a separate event. None should be labeled as the other. Federal Reserve monetary-policy calendar
The test is what happens after the purchase order. Capacity must become usable, faster tasks must become reliable completed work, and the resulting value must reach someone as income, profit or a better product. Following those steps offers a clearer picture of the AI economy than treating either a spending boom or a successful demonstration as the final result.

Comments
Post a Comment