Why the New Fed Chair Talks Like a Startup Founder

The First Hearing Where Optimism Preceded the Data

On July 14, Kevin Warsh, the newly installed Fed chair, told Congress at his first hearing after taking office: “AI is probably the most consequential change to hit our economy since I’ve been an adult.” He said this alongside a pledge to bring inflation to an end.

This will sound familiar if you’ve been reading along. I’ve already covered, twice in this newsletter, the gap between investment measured in the ₩1,000 trillion (~$700 billion) range and economic indicators that barely register it. In We poured in $250 billion — so why isn’t it showing up in the economic data?, I looked at the time lag typical of general-purpose technologies. In The office automation era, 47 years ago, I dug into the gap between a 3x perceived leap and a mere 1.8% measured gain. The FT report cited this time puts AI-related investment at $725 billion — a figure with a different scope and timeframe than the $250 billion from that earlier issue, so the two shouldn’t be conflated.

Today, instead of asking yet again “where’s the productivity,” I want to look at what decisions this unconfirmed productivity is already being used to justify. AI’s productivity gains haven’t shown up in economic indicators yet — but they’re already being invoked as grounds for policy. The first case in point: U.S. monetary policy.

Three Pieces of Evidence Against ‘It’ll Come If We Wait’

My last two pieces landed on the side of waiting a bit longer. General-purpose technologies take time to show their effects, so the argument was: wait until organizations redesign how they work. But over just the past month, three separate pieces of evidence have surfaced against that “it’ll come if we wait” hypothesis.

First, the depth of adoption has stalled. Look at the Real-Time Population Survey (RPS), run by economists at the St. Louis Fed: as of Q2 2026, 45% of the U.S. working-age population uses generative AI at work at least once a week. But the share using it daily has been stuck at around one in ten for over a year now. Even among users, daily usage time is only about 30 minutes, and the time savings that figure implies come out to roughly 10 minutes a day. The lag hypothesis rests on the premise that usage keeps deepening — but while the number of people who’ve tried it keeps growing, the share using it every day for actual work hasn’t moved. We need to distinguish between broadening exposure and deepening use in daily work.

Second, even the productivity gains we’ve already seen may not belong to AI. Chairman Warsh points to the fact that structural productivity¹ growth over the past four quarters has run in the high-2% range as grounds for optimism. But when Barclays economists adjust that figure for cyclical capacity utilization, it drops to the low-1% range. Their explanation: during the post-pandemic period of labor hoarding² (2022–24), when companies kept holding onto workers even as demand cooled, productivity readings understated firms’ true capability — and now, as that excess labor gets shed, the statistics are mechanically bouncing back. In other words, this isn’t a technological leap; it’s a cyclical reversal.

Third, evidence has started emerging that goes beyond “we just haven’t seen it yet” to something starker: industries that adopted AI early aren’t showing different productivity gains either. If AI were truly lifting productivity, we should see it first in fast-adopting sectors like finance, IT, and consulting. But when Barclays ran the correlation between industry-level adoption rates and productivity gains through multiple specifications, the result was “statistically indistinguishable from zero.” A Fed economists’ note released on the 17th shows the same picture: grouping industries into high, medium, and low AI-adoption tiers, the productivity trends across all three groups tracked each other closely over time.

None of this means the lag hypothesis is wrong. But it does mean the optimism that the effects are “just about to show up” deserves a second look.

Unconfirmed productivity is already being used to justify rate cuts

Here’s the thing: this unconfirmed productivity is already being used as a policy justification. The target is monetary policy.

Let me lay out Chair Warsh’s remarks in chronological order. In a Wall Street Journal op-ed last November, before he took office, he wrote that “AI will be a significant disinflationary force” and that “a 1-percentage-point rise in annual productivity growth doubles living standards within a generation.” Earlier this month, at a European Central Bank forum, he cited structural productivity in the high-2% range as grounds for “reason to be optimistic.” And at this hearing, he said AI appears to be raising productivity without displacing workers—tying the promise of ending inflation to the direction of rate cuts, all in a single sentence.

The logic is simple. If productivity rises, the economy can grow without price pressure, which means it’s safe to cut rates. If productivity had actually risen, this would be textbook-correct. The problem, as we just saw, is that the premise isn’t in the data yet.

The FT points out that this narrative resembles the logic once used to explain quantitative easing³: a promise to grow the economy without stirring up workers’ wage-bargaining power. The difference is that QE’s scale was recorded in trillion-dollar figures on the Fed’s balance sheet, while the AI productivity narrative isn’t recorded anywhere. Expectations about AI aren’t a policy tool like asset purchases, whose scale gets logged in the Fed’s books. And there’s no more convenient justification for either party—a White House that has openly wanted rate cuts, or a newly installed chair who took office promising a policy “regime change.”

Even within the same institution, opinions diverge. Fed staff economists released a note on July 17th that withholds judgment, saying “there’s no visible difference in productivity trends across industries”—while the head of that same institution speaks at a hearing about policy direction as if optimism were already the premise. The research division says it’s not confirmed yet; the policy statements speak as if it already were.

Another feature of this narrative is that, by design, it’s hard to falsify. The line “it’s not in the statistics yet, but it’s coming” survives no matter what data comes out. According to the FT, media citations of Solow’s productivity paradox are on pace to set a record high for 2 years running—and I read that frequency as a signal that the “not yet, but coming soon” explanation is needed more and more often. When a debate from 40 years ago gets cited again, it means a lot of people are looking for grounds to defer judgment right now.

What Happens When You Cut Rates on an Unconfirmed Assumption

Let’s start with the question of sequence. There’s a difference between cutting rates after confirming a productivity gain, and cutting rates by assuming one. Neither path escapes uncertainty, but the more you assume a productivity gain that hasn’t yet been confirmed, the more policy comes to depend on forecast. If the assumption holds, it’s a vindicated call. If it doesn’t, the burden lands back on prices. That’s why Barclays bothered to insert this line into its report’s conclusion: “We would be against easing monetary policy on this basis.”

There’s also a mismatch between the timeframe over which change actually appears and the tenure of the policymaker making the call. If AI is, as Fed Chair Warsh put it, “the most consequential change,” then history tells us the timeline for such transitions runs in decades. These kinds of shifts have also been chances for latecomers to catch up with incumbents — that’s what happened with manufacturing automation, and with online retail. A Fed chair’s term, meanwhile, is 4 years. If you build policy on the assumption that gains which will only materialize over decades will show up within a 4-year term, whatever gap opens up between assumption and reality has to be absorbed somewhere else. And costs pile up even while the optimism holds. Asset prices rise on the back of that optimism, data centers get built, hiring and budgets get reallocated. When the outlook gets revised, it’s not just words that change — the asset prices that rose, the facilities that got built, and the people and budgets that got shifted all have to adjust together.

And this isn’t just an American story. Sentences that begin with “since AI will boost productivity” are multiplying fast in Korean policy statements and corporate announcements too. The sentence itself isn’t the problem. What matters is the decision that follows it — which budget gets drawn up, which reorganization gets carried out, which rate decisions and investments get approved on the strength of that explanation. And who ends up bearing the burden when the assumption turns out to be wrong.

Oswarld’s Lens

Honestly, this rhetoric wasn’t unfamiliar to me. It’s the language of startups.

Doing GTM strategy consulting, I’ve run into the sentence “the metrics aren’t there yet, but they’re coming” more times than I can count. Companies without a track record raise money on a story about the future instead of results, and that line is almost never missing from the pitch. What’s striking is that the same rhetoric is now coming from the institution that’s supposed to be the most careful of all. When a startup’s forecast is wrong, investors and employees bear the loss. When a central bank loses the same bet, the cost gets spread across the entire population in the form of inflation.

Here’s one more thing from my experience with data: the stronger the narrative, the more the numbers tend to get measured to fit it. When the U.S. Census Bureau broadened the wording of its question on AI usage, the index jumped by nearly 10 points, and one domestic Korean survey that sampled only companies’ AI/IT staff came up with a usage rate of 55.7%. The U.S. survey’s 21%, representing businesses as a whole, and the Korean survey’s 55.7%, drawn from people whose job is AI, aren’t lies — either one. But looking at which of the two numbers gets put front and center tells you something about who needs which explanation right now.

Let me leave this here for balance: not showing up in the statistics isn’t proof that something isn’t working. I’ve seen this pattern countless times in tech markets — the forecast turns out right, only the timing and the path were wrong — and AI productivity will eventually show up too. My point today isn’t that it won’t come. It’s the fact that the central bank is, right now, betting that it will come within this rate cycle — that’s the point itself.

Closing

The evidence keeps piling up against the idea that AI productivity gains are just around the corner. The share of people using AI daily hasn’t grown in over a year. The recent productivity uptick shrinks to the low 1% range once you adjust for utilization rates. And the correlation between industry-level adoption and productivity improvement is statistically indistinguishable from zero. Meanwhile, this still-unconfirmed productivity boost is already being cited as grounds for interest rate cuts. So what deserves scrutiny right now isn’t the productivity statistics themselves, but who is making what decisions on the basis of that narrative.

Next time you hear someone invoke “AI productivity,” check two things. What is the speaker using that narrative to justify right now? And is there a falsification condition—some piece of data that would make them abandon the forecast? If no conceivable data would change their mind, the claim isn’t much use as a basis for judgment.

Has your own organization ever had a moment where “the numbers aren’t in yet, but they’re coming” won out over actual measurement? Tell me in the comments what decision that narrative pushed through, and how it turned out. If enough examples come in, I’ll dedicate a future issue to “moments when narrative beats data.”

---

📨 If you know a colleague who has to make calls caught between AI investment hype and policy rhetoric, please share this piece with them.

---

Looking at the fragment, I need to check for glossary compliance, number accuracy, and structural fidelity.

References & Further Reading

Primary sources

• FT Alphaville, “Is AI productivity growth in the room with us right now?”, Financial Times, 2026.7. Link ··· This is where today’s piece starts. The key chart and argument from the non-public Barclays report are laid out here.
• Soto, P., Thieu, M. & Allen, J., “The AI Buildout and the Economy: Publicly Available Data to Assess AI’s Impact”, FEDS Notes, Federal Reserve Board, 2026.7.17. Link ··· The heart of this note is a comparison of productivity trends across high, medium, and low-adoption industries. It uses only public data, so you can replicate it yourself.
• Bick, A., Blandin, A. & Deming, D., “The Rapid Adoption of Generative AI”, NBER Working Paper 32966, 2024 (updated quarterly). Link ··· This is the methodology paper behind the Real-time Population Survey (RPS), the source of the adoption-rate and usage-hours figures.
• Warsh, K., “The Federal Reserve’s Broken Leadership”, The Wall Street Journal, 2025.11. Link ··· This op-ed lays out Warsh’s thinking most clearly, from before he took office. This is where the “disinflationary force” quote comes from.
• FedScoop, “Fed chair says AI ‘hasn’t displaced workers’ so far, has boosted productivity”, 2026.7. Link ··· This article summarizes his remarks at the July 14 House hearing.
• U.S. Census Bureau, “Business Trends and Outlook Survey”, 2026. Link ··· This is the source for the 21% corporate adoption rate and the story about the change in question wording.
• CIO Korea, “In 2026, 85% of Korean companies adopt generative AI — 8 in 10 expanding budgets”, 2026. Link ··· This is a Megazone Cloud/Foundry survey of 749 companies in Korea. I’d recommend placing it next to the U.S. statistics to see how differently the two surveys are designed.

Background

• “Productivity paradox”, Wikipedia. Link ··· This lays out the 40-year debate around Solow’s paradox. Seeing how it was resolved in the 1990s gives the word “not yet” a different weight.

Worth reading alongside this issue

• $250 billion poured in — so why doesn’t it show up in the economic data? ··· This is Part 1 of today’s piece. It’s where I laid out the lag hypothesis for general-purpose technologies.
• History repeats itself? 47 years ago, the 📠 office automation era ··· This piece diagnosed the gap between “it feels like 3x” and “measured at 1.8%” as an organizational problem. It pairs well with today’s new evidence.

---

The author is Oswarld (Kwangseob Ahn). Current roles: Adjunct Professor at Sejong University, Strategy Consultant at INLEVEL9. Career, research, books, and recent work are kept current on the About page. Latest · July 2026: HEMA-2: A Consolidation-Aware Tri-Memory Architecture with Multi-Channel Scheduling for Lifelong Conversational AI.

📝 Glossary

Footnotes

1. Structural productivity: The rate of productivity growth once you strip out temporary swings from booms and busts — the economy’s baseline fitness, so to speak. It’s used as a gauge for potential growth. ↩

2. Labor hoarding: When companies hold onto workers instead of laying them off, even as demand falls. This happens when rehiring is difficult or hiring costs are high. During these periods, output falls with the same headcount, which drags down measured productivity. ↩

3. QE (Quantitative Easing): A policy in which the central bank buys up large quantities of assets, such as government bonds, to inject money into the market. It’s used as a stimulus tool when rates can’t be cut further — and its scale shows up directly on the central bank’s balance sheet. ↩