Stanford expert: The U.S. is entering the "AI harvest period," with productivity growth expected to double to 2.7% by 2025

The British Financial Times recently published a commentary article with a straightforward theme: The productivity "takeoff" brought by AI may finally be visible in macro statistics.

The author of the article is Erik Brynjolfsson, the director of the Digital Economy Lab at Stanford University and a co-founder of a company called Workhelix that studies AI and organizational efficiency. He stands at the forefront of academic research while also observing the real-world implementation of AI in businesses.

In this article, he presents a core judgment: the United States may be transitioning from the "AI investment phase" to the "AI harvest phase." He cites the latest economic data to indicate that the previous situation of "talking about AI everywhere, but not seeing it in productivity data" is changing.

More specifically, based on updated data, he predicts that the productivity growth rate in the U.S. will reach approximately 2.7% by 2025, nearly double the average level of 1.4% over the past decade. If this trend solidifies, it means that AI is no longer just a story in PowerPoint presentations but is beginning to translate into measurable efficiency gains in GDP.

Signals from Macro Data: Output Hasn't Dropped, but Employment Has Decreased

Brynjolfsson starts with a "counterintuitive" macro adjustment: the benchmark revision by the U.S. Bureau of Labor Statistics shows that total payroll employment growth has been revised down by about 403,000 jobs. At the same time, U.S. economic output has not weakened, with real GDP remaining strong, growing at 3.7% in the fourth quarter.

He describes this combination of "high output with significantly lower labor input" as a typical characteristic of productivity growth, directly stating: “This decoupling — maintaining high output with significantly lower labour input — is the hallmark of productivity growth.” In other words, the same or even more work is being done with fewer people, naturally leading to increased productivity.

However, the author also cautions against excessive excitement, as productivity data is inherently "volatile," and short-term readings are easily influenced by statistical revisions and cyclical factors, "requiring more periods for validation."

In this regard, MIT economist Daron Acemoglu has previously pointed out in public research that the impact of AI on overall productivity depends on whether it can genuinely replace or enhance labor in a sufficient number of tasks, rather than just providing "localized efficiency improvements."

The "J-Curve" Explains Why It Is Only Now Becoming Apparent: First Plant Trees, Then Bear Fruit

Brynjolfsson places the diffusion path of AI within a longer historical framework of technology. He mentions that the economics community has long been entangled in a modern version of the "Solow Paradox."

He summarizes this awkwardness with a quote: “we have seen artificial intelligence everywhere except in the productivity statistics.” In other words: AI is everywhere, but productivity in statistics remains stagnant.

The explanation given is the "productivity J-curve." Many general-purpose technologies, from steam engines to computers, do not immediately boost productivity upon installation, but rather go through an "investment period."

Companies need to reorganize processes, train employees, and redo business models, much of which involves intangible capital and may temporarily suppress "measurable productivity." Only after the organizational transformation is complete does the "harvest period" begin, and efficiency is reflected in the data.

Economic historian Paul David studied the "productivity lag" during the electrification era and found that after factories switched from steam power to electric motors, significant efficiency gains often only appeared after the layout, process organization, and management methods were all restructured. The "first transform the organization, then see statistical returns" that AI is currently facing is essentially the same logic.

Micro-level changes are happening: entry-level hiring down 16%, but "power users" are compressing timelines

In addition to macro data, the author provides micro evidence. His research with collaborators Bharat Chandar and Ruyu Chen found that in industries with high "AI exposure," entry-level job recruitment has noticeably cooled, with hiring for junior positions dropping by about 16%.

On the other hand, those who use AI to enhance their skills are seeing job growth. The author's interpretation is that companies have begun to use AI for some "codifiable and standardized" entry-level tasks.

The author also distinguishes between "potential" and "realized benefits." Many companies currently only use generative AI for lightweight scenarios like translation and summarization, describing this usage with a sharp term: "glorified dictionary."

However, among a small group of "power users" observed in his company, AI agents can automate end-to-end processes through interactive dialogue, such as directly generating complete marketing plans, compressing "weeks of work into a few hours." He emphasizes that the real challenge for companies lies not just in acquiring technology, but in how to apply it.

From external research, this phenomenon of "a few people reaping the benefits first" is not surprising. McKinsey has emphasized in multiple industry reports that the value release of generative AI largely depends on process reengineering and personnel retraining, rather than simply buying tools.

From experimentation to structural utility: the next competition will be about organizational capability and macro environment

At the end, Brynjolfsson provides a stronger trend judgment: “We are transitioning from an era of AI experimentation to one of structural utility.” This means that AI is moving from a phase of "let's try it out and play with it" to a stage where it can consistently provide value. For companies, the next competitive focus will shift from "whether there is a model" to "whether the model can be embedded into the business framework." What should companies specifically do? Based on the author's viewpoint, it can be summarized into three things: First, do not stay at the "advanced dictionary" level; embed AI into end-to-end processes and let it participate in delivery rather than just serving as an assistant; Second, upgrade training goals from "being able to use tools" to "being able to use AI to redo work methods," allowing the average employee's capabilities to be enhanced by AI; Third, track returns with data and metrics to avoid the difficulty of reviewing and scaling after a period of excitement.

At the same time, he also reminds that external risks may offset efficiency gains, including "geopolitical trade wars" and macro headwinds such as misjudgments in fiscal and monetary policies. From a neutral perspective, technological advancement and macro governance are two separate lines: the former provides possibilities, while the latter determines whether these can be smoothly realized.