The landscape of the AI revolution can be confusing. Sometimes it appears as if the latest artificial intelligence models offer significant improvements in functionality. One example is when two OpenAI models broke out of a lab, accessed the internet, and compromised another AI company. or as wall street journal The headline read, “The day the bots went wild.”
But then you might come across another headline. wall street journal: “Ignoring predictions of AI annihilation, big companies are starting to hire again.” And from that piece:
The move to at least modestly expand the workforce is a reversal from the corporate messaging that was common for much of the AI era. Large employers have held back on adding staff, largely due to economic uncertainty and the belief that artificial intelligence could take on more jobs. But some executives say the cost and limitations of AI will necessitate adding more talent. Some companies want to rehire people after layoffs.
One thing these two new stories illustrate is the gap between lab capabilities and workplace productivity. This is a disconnect that should come as no surprise to students of economic history. It takes time for companies to reorganize and fully and productively integrate important new technologies into their workflows and operations.
And even though it’s happening faster than previous major technological advances, it remains a work in progress. As Ernie Tedeschi, chief economist at payment processor Stripe, explained in a recent analysis, if such consolidation were to occur on a large scale, we would see things that are currently invisible.
Tedeschi writes:
For example, significant increases in microproductivity are expected to show up in total factor productivity (TFP) estimates. TFP is formally the portion of productivity growth that cannot be explained by increases in the amount of capital or labor input into the economy. More informally, TFP is what economists think of as “pure” technology and efficiency growth, with the important caveat that TFP cannot be directly observed and must be estimated using economic models. If AI is increasing worker productivity on the front lines, then TFP estimates should accelerate as well as overall labor productivity. However, this is not actually the case. Despite strong overall labor productivity numbers, the San Francisco Fed estimates that TFP growth has been near zero over the past year.
(As Tedeschi points out, the Bureau of Labor Statistics puts the TFP growth rate in 2025 at 0.8%, not zero, but that’s the same as the bureau’s 10-year average, so it doesn’t seem like anything special is going on.)
Despite (a) many studies showing that AI can improve worker productivity on certain tasks, and (b) overall labor productivity increases (especially as companies work harder with existing capital to meet A), (c) we have yet to see widespread increases in economy-wide TFP or underlying efficiency from AI. And the final question is the one that will determine the future of the U.S. standard of living, and it’s all the matches that will determine whether all the investments in AI infrastructure will pay off as many companies and investors expect.
Ultimately, this is not surprising considering we are still in the early stages.
But what the current numbers don’t yet bear out is that this transmission [from micro impacts to macro impacts] has already made considerable progress on an overall level. Implementation is too thin, the cross-sectional signal is too weak, downstream bottlenecks remain, and productivity gains are too dependent on utilization to give us confidence in that conclusion. Rather, the acceleration we have seen so far is primarily the result of companies seeking to meet the demand for AI capabilities by pushing the limits of their existing infrastructure.
