90% of firms Moody's surveyed say AI has had zero impact on their own employment or productivity over the last three years — even though 70% of them are already using it.
Job creation has slowed across every industry since ChatGPT's release, Gabriel Agostini says, and it has slowed only a little faster where AI adoption is highest — evidence, he argues, that the labor market's real culprit is the Fed's rate hikes and a post-pandemic hiring correction, not artificial intelligence.
"Yeah, I mean this is the $1.5 trillion question."
Agostini is an Assistant Vice President for Credit Strategy and Standards at Moody's Ratings, and he wrote the report the episode is built around, comparing AI's adoption rate against its measurable effect on jobs and output across the US, UK, Germany and Australia.
I listened to the full interview so you can skip it.
Here are the 7 takeaways that matter.
👤 Guest: Gabriel Agostini, Assistant Vice President for Credit Strategy and Standards at Moody's Ratings, who wrote the report the episode is built around
🎙️ Host: Greg Sobel, Vice President at Moody's Ratings and host of Credit Currents
📰 Published: 9 September 2026 on YouTube
🔴 YouTube | 🟣 Apple Podcasts | ⏱️ 14 min
Key Takeaways
90% of firms report zero measurable impact from AI on their own employment or productivity, even as adoption keeps climbing
US worker AI usage is up to about 45% from roughly a third two years ago
Job creation has slowed everywhere, not just in the industries adopting AI fastest
The New York Fed traces two-thirds of the rise in youth unemployment to remote work making it harder to mentor junior staff, not to AI
AI's own footprint in the economy is still narrow — Anthropic's usage data shows just 16% of occupational tasks touched by Claude conversations
Productivity dips before it rises, the J-curve pattern Agostini says every general-purpose technology goes through
The computer took 20 to 30 years to turn into measurable productivity gains; Agostini expects AI's gap to be shorter, not zero
Hyperscalers need to generate $300 billion to $600 billion a year in new economic value just to break even on their AI spending, and the data isn't there yet
1. The Adoption-Impact Gap
Sobel opened by putting the obvious question to his colleague: if AI is as transformational as people think, shouldn't its effects already be visible in the labor market?
Adoption has outrun any measurable effect, in every direction Agostini checked. "Not yet. And the jury's still out over the next few years." Across the US, UK, Germany and Australia, about 70% of firms are already using AI in some form — report generation, visual content, coding — and 45% of US workers report using it for work as of this May, up from about a third two years earlier
Despite that, roughly 90% of the firms surveyed say they have experienced zero impact on their own employment or productivity over the last three years
The labor-market data backs up the survey answers. If AI were displacing workers through automation, Agostini said, the effect should show up fastest in the industries adopting it fastest — what his report calls the pioneer industries, against the laggards where adoption has been slow. Comparing job creation since ChatGPT's rollout, he found it has fallen everywhere, pioneers included, with only a slightly bigger dip where adoption is highest
His read is that broader forces, not AI, explain most of that dip. "We think that bigger macro structural forces explain most of the dip in job creation for those pioneer industries rather than AI itself."
2. Blame the Macro, Not AI
Sobel pushed back with the thing he hears constantly: that AI is the culprit behind the slowdown in hiring. Agostini said the timing makes that an easy but incomplete story.
Three macro forces, in his account, do most of the work the AI story gets credit for. The first is a hiring correction: pioneer industries over-hired for a remote, digital-first future right after the pandemic that didn't fully arrive, and are now adjusting
The second is remote work itself, and he cited a specific study on it. "The New York Fed recently estimated that two-thirds of the rise in youth unemployment, which we usually attribute to AI, is in fact due to remote work making it harder to mentor and train younger employees." Businesses are responding by favoring more experienced hires when a role is remote
The third is monetary policy. "And lastly, we can't forget that in the past few years we've gone through the tightest monetary hiking cycle in 40 years." Agostini said Moody's has empirical evidence that automatable work is more sensitive to rate hikes than other work, making it look like an AI effect when the mechanism is the cost of credit
3. Why Productivity Lags
Sobel asked the natural follow-up: if usage is rising this fast, why hasn't productivity grown in lockstep? Agostini, who said he has seen his own productivity rise, gave three reasons.
AI's usage remains narrow, so its economic footprint is narrow too. It is concentrated in a specific set of cognitive tasks, particularly computer and mathematical roles. "And in fact, if we use Claude's, or Anthropic's rather, own Claude usage data, we find that only about 16% of occupational tasks have been covered by Claude conversations as of November 2025."
Adopting the technology costs more than it returns at first. Reorganizing workflows, building data infrastructure and retraining staff all come before any payoff, which produces a period of lower, not higher, productivity. "That's why economists tend to call that pattern a productivity J curve, because you go down before you go up."
The gains that exist are still individual, not organizational. People are writing emails and summarizing reports faster, in his telling, but that has not yet aggregated into a firm-wide or economy-wide number. "It's too early in the process for this individual improvement to translate into a firm-wide, economy-wide improvement."
4. The 20-Year IT Precedent
Sobel noted that Moody's has previously focused on which jobs are exposed to AI, while this report instead measures actual usage and adoption — and asked what that shift shows. Agostini pointed to the personal computer.
The first commercial microchip predates the productivity payoff by decades. "For context, the first microchip for commercially available purposes was released in 1971 by Intel."
It took roughly 15 more years for computers to become a normal part of the workplace, and another 10 years after that — into the mid-to-late 1990s — before the technology showed up in higher productivity growth
He does not expect AI to repeat that full timeline, but he does expect a real gap. "I don't think we're on a 20 to 30-year timeline when it comes to AI, as we've said its adoption and usage is moving far faster, but I think the precedent does suggest to us that we should expect some sort of gap between the introduction of the technology and our ability to turn that technology into improved productivity."
5. The Investment's Breakeven
Sobel turned to the money: businesses and markets are investing on the expectation that productivity gains arrive on a specific schedule. He asked what happens if they don't.
He sized the bet first. "Yeah, I mean this is the $1.5 trillion question." That is what he says hyperscalers — "the Googles, the Metas" — have invested since 2023 in the AI buildout: data centers, chips and IT infrastructure
The breakeven math is specific. By his calculation, that investment needs to generate $300 billion to $600 billion a year in new economic value to cover its funding costs and depreciation — through improved business workflows, more efficient consumer search, or entirely new markets such as AI-enabled drug discovery
In growth terms, that translates to productivity growth of 0.2 to 0.5 percentage points over the next few years, rising as high as 0.8 points if hyperscalers add another $1 trillion of capex in 2027, which Agostini expects them to
Executives think they'll clear the bar; the data doesn't show it yet. Surveyed executives expect productivity growth of about 0.8 percentage points a year for the next three years — enough to break even — but Agostini said that gain has not shown up in the data so far, which is the gap between expectation and reality
A precedent exists for the scale of gain needed, but the size of the jump from here is what worries him. The 1990s computer and dot-com rollout produced a comparable rise in productivity, so it "wouldn't be unprecedented" — but getting there from where the data sits today would be a step change, not a continuation
He named the downside directly. "If we don't get those productivity gains in either scale or speed, we would expect to see a pretty meaningful disruption to economic activity and some indigestion, quite a bit of it within the AI ecosystem." Put together, he said, that adds up to a meaningful economic slowdown: "In the early 2000s, we got a recession because of it."
6. What Investors Should Watch
Sobel asked what indicators he'd be watching over the coming months.
Quarterly productivity growth is the number that settles the question. "Businesses are adopting AI because they think it's going to improve productivity and boost profits. The hyperscalers are investing in it because they expect the same, right?" — making productivity data the test of whether that bet is paying off
Hyperscaler capex itself is the second signal. "They've announced to the markets that they're basically willing, ready, and able to invest around $1 trillion next year in the AI build out, which is a good signal of confidence in the technology itself and what they expect to generate in returns." A pullback from that figure, he said, would be a sign hyperscalers themselves are losing confidence in the near-term return
7. The Dot-Com Silver Lining
Sobel closed by asking about the parallel everyone draws to the late-1990s dot-com boom — and whether, if AI is on a similar path, there's a silver lining to look forward to.
He leans on the parallel deliberately, because the underlying pattern matches. Dot-com was also a technology built on hard and soft infrastructure that changed how work got done, which is why he thinks the precedent holds
The infrastructure outlasted the bust, in his account, and did the compounding later. "Although financial markets adjusted their value of what they thought that technology was worth, under the hood, the hard and soft infrastructure was still being built, still being developed." "That stuff remained with us, is still with us today, and drove the next 10, 20, 30 years of growth."
He expects the same pattern from today's frontier models, regardless of what happens to AI-company valuations or investor expectations of profitability between now and then
Bonus Insights
Sobel's summary of the stakes drew a one-line answer back. "A lot of heartburn for a lot of people." "Yeah, simply put."
Sobel's closing reaction to the dot-com parallel was short. "Okay, that's comforting." Agostini: "Yeah, yes, it should be."
Agostini's bottom line is that the absence of an AI productivity boom in today's data isn't evidence the technology won't pay off — it's the same lag every prior general-purpose technology showed before it did.
Products, Companies & Tools Mentioned
Intel (Released the first commercially available microchip in 1971 — Agostini's reference point for how long the computer took to show up in productivity data)
Anthropic's Claude (Its own usage data is the source for Agostini's figure that only 16% of occupational tasks have been touched by Claude conversations)
Google and Meta (Named by Agostini as examples of the hyperscalers behind the roughly $1.5 trillion in AI buildout capex since 2023)
ChatGPT (Its rollout is the reference date Agostini's report uses to compare job creation in AI pioneer industries against laggards)
Books & Resources Mentioned
Widespread AI adoption is yet to affect labor market or productivity – Gabriel Agostini, Moody's Ratings (The report the whole episode is built around)
Hyperscalers' asset-heavy model spurs borrowing, equity sales – Moody's Ratings (Related research on the hyperscaler capex Agostini sizes at $1.5 trillion since 2023)
If this was worth your time, send it to someone who has to have a view on this.
Get the latest market chatter as it happens: