Context

This is a Rapid Response segment inside the Masters of Scale feed, hosted by Bob Safian. His guest, Austan Goolsbee, president of the Federal Reserve Bank of Chicago, joins ahead of the Fed's Jackson Hole meeting to talk through tariffs, inflation, and how the Fed operates under new chair Kevin Warsh. It's a macroeconomics conversation, not a product one, but it's relevant to a PM for two reasons: it's a real-time case study in making high-stakes decisions under genuine uncertainty, and it includes a candid, skeptical read on AI's actual economic impact versus its hype, from someone with no product to sell.

The Big Idea

Good decision-making under real uncertainty depends on deliberately building in diverse perspectives, choosing metrics that won't mislead you when the underlying population shifts, and resisting the pull to treat hype as evidence, whether the subject is inflation or AI.

Goolsbee's account of how the Fed operates, and his skepticism about AI's current economic impact, both come back to the same discipline: separate what you can actually measure and verify from what everyone around you is excited or panicked about.

Key Insights

Diverse committees catch individual blind spots

The Federal Open Market Committee that sets interest rates has 19 people: 12 regional reserve bank presidents and 7 political appointees on staggered 14-year terms, each with a different professional background and worldview. Goolsbee argues this structure matters most exactly when uncertainty is highest, since it means no single person's read on the economy goes unchallenged. "My colleagues can change my views or how I interpret the data," he said, describing it as one of the most important features of how the committee is set up, not a bureaucratic inconvenience.

Forward guidance trades clarity for flexibility

New Fed chair Kevin Warsh dislikes "forward guidance," the practice of publicly committing to a rule like "if X happens, we will cut rates." Goolsbee laid out the trade-off directly: explicit guidance gives markets and the public clarity, but it also ties the Fed's hands to a promise made under yesterday's information. Give too little guidance instead, and people fill the silence with their own assumptions, which can add more volatility, not less. Under Warsh, the Fed's post-meeting statements have already gotten shorter and dropped most of that forward-looking language, a real, visible shift in strategy, not just a stated intention.

Population shifts can quietly break a metric

For years, Goolsbee treated the monthly jobs-created number as one of the single most informative economic indicators available. During the recent immigration crackdown, that number fell sharply, and many observers read it as an early recession signal. Goolsbee's response was to stop trusting it: when the number of people available to work is itself changing, a falling jobs-created count doesn't tell you whether the economy is weakening or simply has fewer people left to hire. He switched instead to rate-based measures (the unemployment rate, the vacancy rate, the hiring rate), which stayed stable through the same period and gave a truer read.

AI hype is outrunning delivered productivity, for now

Goolsbee has watched productivity growth pick up over the past two and a half years, loosely tied to AI and machine learning adoption, then flatten over the last six months despite continued heavy investment. He's candid that the sectors seeing "high productivity immediate" gains right now are narrow, and that many companies report spending heavily on AI without yet finding the productivity payoff. His read, informed by his own research into the dot-com era, is that the underlying technology can still be transformative, "the internet did change the whole world," while the near-term timeline the most bullish voices predict is usually wrong. He's blunt about the fiscal risk this creates too: if companies keep spending trillions on data centers premised on future productivity that hasn't shown up yet, that speculative spending can overheat the economy in the short run, independent of whether the long-term bet turns out to be right.

AI helps rules-based, information-heavy work first

Asked where AI is actually delivering today, Goolsbee pointed to a specific example: the head of a major arbitration agency told him AI is already as good as or better than human arbitrators, because a case's facts and evidence can be fed in directly and processed against clear rules. By contrast, in manufacturing, the district with the heaviest concentration of factories in the Fed system is "still looking for the use cases," since AI's biggest potential impact there depends on robotics and physical automation that hasn't matured yet. The pattern: adoption and payoff show up first where the task is already structured, rules-based, and information-dense, not where it requires physical action in the world.

Mental Models & Frameworks

The lump of labor fallacy

A recurring argument Goolsbee pushes back on: the idea that there's a fixed "lump" of work to be done, so any task AI can do better than a human permanently removes that person from the workforce with no new work to replace it. He notes this exact argument has been made about earlier waves of automation and technology, and been wrong every time, without claiming it's guaranteed to be wrong again. Use it as a check whenever a claim about AI (or any new technology) assumes total job displacement with no offsetting new demand: ask what evidence supports this time being structurally different from the last several times the same argument was made and failed.

Principle: verify a metric's population before trusting it

When: a headline number (a count, not a rate) moves sharply and gets treated as a clear signal of underlying strength or weakness. Why: a raw count conflates two different things, the underlying rate of change and the size of the population it's counted over, so if the population itself just shifted (fewer workers available to hire, fewer customers eligible to buy), the raw number will mislead even though nothing about the underlying behavior changed. Goolsbee's fix was switching from the monthly jobs-created count to rate-based measures once he suspected the labor supply itself, not labor demand, had moved.

Practical Application

Audit a metric before trusting a swing in it

Before reacting to a sharp move in any headline number your team tracks (signups, churn, usage), check whether the underlying population changed at the same time, a marketing push that shifted who's in the funnel, a pricing change that shifted who's eligible, a policy change that shifted who's active. If the population moved, look at a rate or ratio instead of the raw count before deciding what the number means.

Build review panels with genuinely different vantage points

Before a high-stakes, hard-to-reverse product decision, deliberately assemble reviewers who won't naturally agree with each other, not just a bigger group of similar people. The Fed's committee works because a market economist, a former Council of Economic Advisers chair, and business-background regional presidents are structurally required to reconcile different reads on the same data before acting.

Sort AI opportunities by how rules-based the task is

When evaluating where to invest in AI features or workflows, prioritize tasks that are already information-dense and governed by clear rules (document review, structured decision support) over tasks that require physical action or judgment calls with no clear rule set. That's where Goolsbee sees real productivity gains landing today, while physical-world use cases are still searching for their first proven application.

Separate hype-driven investment from proven payoff

Before committing significant budget to a new technology on the strength of its long-term promise, track whether the investment is being justified by a measured productivity gain you can point to, or by the expectation that gains will show up eventually. Goolsbee's warning about AI applies just as directly to any internal bet: heavy spending premised on future payoff that hasn't materialized yet is a real risk in itself, regardless of whether the long-term bet is ultimately right.

Questions to Consider

• Is there a metric on your team's dashboard right now whose underlying population (who's eligible, who's included, who's active) has shifted recently in a way that would make a swing in the raw number misleading?
• When your team debates whether to commit to a specific public roadmap promise versus staying vague, are you weighing the credibility gained from clarity against the flexibility lost from being locked in, the way the Fed weighs forward guidance?
• If someone on your team argued that a new AI capability will make an entire role or workflow obsolete with nothing to replace it, what evidence would you ask for before accepting that this time is different from past predictions of the same kind?
• Which of your current AI investments are you justifying by a productivity number you can already measure, and which are you justifying by an expectation that the payoff is still coming?

Bottom Line

Confusing conditions, whether it's an economy sending mixed signals or a new technology surrounded by hype, call for the same discipline: build in genuinely different perspectives before deciding, check whether your metrics are still measuring what you think they're measuring, and separate what's actually been proven from what everyone expects will eventually be true.