Why AI's Productivity Gains Don't Show Up in GDP Yet

The English draft is accurate and complete. No corrections needed.

In the field of GTM (go-to-market) strategy, I’ve watched this scene play out again and again: a company buys a new tool, and work doesn’t get faster the way everyone expected. The same thing happened when organizations adopted SaaS or CRM systems, or migrated to the cloud. Teams would bolt the new tool onto their existing workflow, and only afterward start rethinking approval processes or how departments actually collaborate.

That experience came back to me as I watched the debate over AI’s productivity effects unfold. Individuals report writing document drafts faster than ever, yet whether AI has moved the needle for the economy as a whole is still very much contested. The two claims aren’t necessarily contradictory. How long an individual task takes, how long it takes an organization to finish its work, and productivity across the entire economy are three separate metrics.

More AI investment doesn’t confirm better work efficiency

Investment in building data centers and buying servers expands economic activity. Whether the companies using that equipment actually produce more goods or services in the same amount of time is a separate question that needs its own verification. The dollar figure of AI-related investment alone tells you nothing about how much the productivity of workers actually using AI has improved.

Productivity statistics themselves also demand careful interpretation. In a February 2026 analysis, Martha Gimbel of Yale’s Budget Lab explained that while U.S. productivity growth over the past several quarters has been relatively strong, it isn’t unusual by pre-pandemic standards. The output and labor-hour data used to calculate productivity are also subject to revision, which makes it hard to draw firm conclusions from early figures alone.

There’s more than one reason an average can rise. For instance, if employment shrinks in industries where measured productivity is low, the overall average can climb even if the remaining workers’ actual way of working hasn’t changed at all. To credit any productivity gain to AI, you have to rule out these kinds of compositional shifts and other factors, like changes in equipment utilization rates.

None of this means AI has no effect. It means we still have to figure out how much of the change showing up in the statistics is actually attributable to AI use. That’s exactly why a good quarter or a bad quarter of numbers isn’t enough to draw a conclusion.

Using AI once isn’t the same as AI being embedded in your workflow

An August 2024 survey introduced by the St. Louis Fed found that about 40% of US respondents aged 18-64 said they’d used generative AI at some point. That figure includes use outside of work entirely. You shouldn’t read this as meaning 40% of workers use AI every day on the job.

Who you ask matters just as much. BCG’s 2025 survey covered 10,635 office workers across 11 countries and regions. Among executives and managers, more than three-quarters said they use AI regularly, but among frontline employees, that figure was 51%. “Regular use” here means daily or several times a week.

A company simply saying it has “adopted AI” tells you little about how deeply it’s actually being used. You need to know which roles, which tasks, and how often. And you can’t assume an executive’s experience with AI mirrors the working conditions of frontline staff.

No Hangul, numbers, and headings all check out — no changes needed.

Cutting writing time by 40% doesn’t make the whole company 40% faster

There is research confirming the effect at the level of individual tasks. In a 2023 MIT study, 453 professionals were given work-related writing tasks and compared based on whether they used ChatGPT. The group that used ChatGPT saw average task time drop by about 40%, and their output was rated about 18% higher in quality.

That said, the study’s tasks didn’t fully reproduce real jobs that require knowledge of internal company context or complex fact-checking. So the results can’t be applied uniformly across every role.

If writing is only part of an employee’s job, a reduction in writing time doesn’t translate into an equivalent reduction in total working hours. Even if a draft gets finished faster, it may still have to wait for review and approval — and you’d need to check whether the time saved can actually be redirected to other productive work.

The time spent checking and revising AI-generated output also has to be counted. If you only measure how long it takes to generate a document, you miss the cost of catching errors or rewriting content. You only get a true picture of how much AI actually helps by measuring both the speed of producing a draft and the speed of completing the final output.

Changing how work gets done takes time, too

This isn’t the first time new technology has spread quickly while its effects stayed hard to spot in the statistics. In 1987, Robert Solow pointed out that even as computers became widespread, productivity statistics failed to show clear gains. This observation became known as “Solow’s productivity paradox.”

Erik Brynjolfsson, Daniel Rock, and Chad Syverson’s research on the “productivity J-curve” focuses on the additional investment required to actually put new technology to use. Beyond simply buying software, resources go into training employees, restructuring organizations, and building new ways of working.

The value of this preparation may not be fully captured by conventional statistics. The explanation goes like this: productivity gains get underestimated in the early stages of adoption, and then the increase looks especially large later, once the effects of that preparation start to show. The J-curve is a concept that explains this kind of measurement shift. It’s not a prediction that every company adopting the technology will inevitably take a loss at first and succeed later.

This concept alone can’t tell us where AI currently sits on that curve. But it does give us reason to look beyond model performance or tool spending, and ask what companies have actually done to prepare for real-world use.

Oswarld’s Lens

When I was building GTM strategies, the problem I kept running into usually wasn’t the tool itself — it was the work process around using it. You roll out a new feature, but if the existing approval chain and division of labor between departments stay exactly the same, the feature never delivers its full potential. Once you start changing how the work actually gets done, though, you’re forced to reconsider what you hand off to the tool and what still needs a human judgment call.

I think this matters just as much with AI. Say drafting got faster — you still need to define what the reviewer checks for, who makes the final call, and how errors get corrected. Rolling out accounts is one task; defining these procedures is another, and both are necessary.

BCG’s survey found that roughly half of respondents’ companies had already moved past simple tool deployment into redesigning their work processes. Respondents at organizations that had undertaken this redesign reported bigger time savings and greater improvements to their work. That said, this is a difference between survey groups — you can’t conclude that redesign alone caused the performance gap.

So when people argue over whether AI actually works, I’d rather start by comparing specific tasks before and after adoption. Not just how many minutes it took the model to produce a draft, but how much faster the whole process got once you include review and approval, whether output quality held up, and where the saved time actually went.

Judging macroeconomic statistics requires more data than we currently have. In the meantime, individual organizations can measure what’s actually observable to them. What matters to me on the ground isn’t the fact that a tool was adopted — it’s whether the people using it can actually get their work done better than before.

References & Further Reading

• Yale Budget Lab, An AI Productivity Boom? Don’t Count Your (Productivity Data) Chickens, 2026.2.19. An analysis of how productivity statistics get revised and interpreted.
• St. Louis Fed, How Much Are Businesses Using Artificial Intelligence?, 2025.2.5. Draws a distinction between individual usage experience and actual corporate AI adoption.
• MIT, Study finds ChatGPT boosts worker productivity for some writing tasks, 2023.7.14. Lays out the results and limitations of a writing-task experiment.
• Brynjolfsson, Rock & Syverson, The Productivity J-Curve, 2021. Examines the relationship between the intangible investment needed to adopt new technology and how productivity gets measured.
• BCG, AI at Work 2025: Momentum Builds, But Gaps Remain, 2025.6.26. Surveys AI usage by seniority level and the redesign of work processes.

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.