Kevin Warsh Poses Nine AI Questions for Federal Reserve Review

Key Takeaways

• Kevin Warsh framed AI as a general purpose technology with uncertain effects on productivity, employment and monetary policy.
• The Federal Reserve’s inquiry focuses on whether AI complements workers, displaces labor or shifts investment toward costly infrastructure.
• Businesses may need better productivity measurements and governance practices before AI spending produces clear economic gains.

Federal Reserve Chairman Kevin Warsh has put artificial intelligence more firmly on the central bank’s policy agenda, raising nine questions about how the technology could affect productivity, labor and capital formation. In remarks at Jackson Hole, Warsh described AI as a “general purpose technology” while emphasizing how little policymakers can confidently model about its eventual economic impact.

The approach is notable for what it does not assume. Warsh did not treat rising AI adoption as automatic evidence of higher productivity, nor did he reduce the discussion to job losses. Instead, his questions covered the relationships among computing demand, investment, employment and economic output. He also pointed to a Federal Reserve task force examining the issue.

As Barron’s reported, the central challenge is asking better questions before folding AI into forecasts and monetary policy decisions. That is harder than it sounds. Central banks rely on historical relationships among wages, output, investment, inflation and employment. A technology capable of changing several of those relationships at once can make past patterns less useful.

One issue concerns tokens, the units processed by generative AI systems. Does greater token consumption complement human work by helping employees complete tasks faster, or does it compete with labor by automating more of the underlying activity? The answer may vary sharply by occupation. Software development, customer support, marketing and research can all use similar models, yet their workflows and productivity measures differ.

A company can deploy an AI assistant and record heavier computing expenses long before it sees measurable gains in revenue or output. Employees need training. Processes need redesign. Data needs cleaning and access controls. Some early benefits may appear as improved quality or faster internal decisions, neither of which is captured neatly by conventional productivity statistics.

That measurement gap matters to the Federal Reserve. An OECD paper published in 2024 concluded that macroeconomic productivity gains from AI are plausible but not yet clearly visible in official data. It also warned that existing statistics may undercount some AI-related improvements. If measured productivity lags behind actual operational gains, policymakers could receive a distorted view of the economy’s capacity.

Capital intensity creates another complication. Current AI development depends heavily on data centers, accelerators, networking equipment and electricity. NVIDIA, Microsoft and Amazon Web Services sit near the center of that investment cycle. But future models could become more efficient, allowing enterprises to run useful systems with less infrastructure. Which path prevails, an escalating race for computing capacity or more capital-light deployment?

The distinction carries consequences for interest rates and inflation. A prolonged infrastructure boom could lift demand for construction, power and specialized hardware. More efficient systems, by contrast, could spread AI capabilities without the same physical investment requirements. Either outcome could raise productivity, but the timing and inflation effects would look different.

Potential value is substantial. McKinsey estimated in 2024 that generative AI could add $2.6 trillion to $4.4 trillion in annual value across global use cases. Capturing that value, however, depends on organizational change, regulation and labor reallocation. Model access alone does not produce the upper end of that range.

For business leaders, Warsh’s inquiry reinforces the case for separating AI activity from AI outcomes. Token volumes, pilot counts and infrastructure spending show adoption, not necessarily productivity. More useful indicators can include cycle times, error rates, customer resolution rates, output per employee and the cost of human review.

Governance also enters the economic equation. The NIST AI Risk Management Framework and the revised 2024 OECD AI Principles offer reference points for accountability, transparency and risk controls. These practices may add near-term costs, but they can also reduce failed deployments and clarify where human oversight remains valuable.

The Federal Reserve is not trying to pick winning models or vendors. Its concern is broader: whether AI changes the economy’s speed limit, the distribution of work and the amount of capital required to generate growth. For now, Warsh’s nine questions signal intellectual caution rather than a settled policy view. That said, caution is not passivity. Better measurement inside enterprises will give both executives and policymakers a clearer picture of whether the AI investment wave is producing durable economic value.