Context
This note follows our prior work on the AI supercycle, AI data-center capex, memory demand, and rate risk. The question here is more basic: is AI actually raising productivity, and is that effect already visible at the macro level?
TL;DR
- The right answer is micro yes, macro not yet fully confirmed.
- Brynjolfsson, Li, and Raymond’s Generative AI at Work finds that AI assistance raised customer-support agents’ productivity by about 15% on average, with the largest benefits for less experienced and lower-skilled workers.
- The Fed’s FEDS Notes shows that AI adoption is no longer trivial: about 18% of firms had adopted AI by year-end 2025, work-related GenAI usage among individuals was around 41%, and employment-weighted survey estimates suggest that 78% of workers were at AI-adopting firms and 54% at firms using LLMs.
- The Kansas City Fed finds that U.S. labor productivity has moved above its pre-pandemic trend since late 2022, but the pickup is not broad-based. AI adoption aligns with faster productivity growth across industries, but explains little of the aggregate shift so far.
- The San Francisco Fed’s policy lesson is that aggregate productivity, labor, and inflation data are not enough. Policymakers and investors need disaggregated evidence and business-level signals before the macro data fully reveal the transformation.
The Four Sources Fit Together
| Source | Layer | Main question | Core answer |
|---|---|---|---|
| Generative AI at Work | Workplace task | Does AI raise productivity on the job? | Yes, in specific workflows |
| Fed FEDS Notes | Adoption | How widely is AI used? | Meaningfully, but unevenly |
| Kansas City Fed | Industry productivity | Is the U.S. productivity pickup AI-driven? | Related, but not fully explained by AI |
| San Francisco Fed | Policy | What should policymakers watch? | Micro and sector data, not only aggregates |
The combined message is straightforward:
AI raises productivity in real workplace settings
→ adoption is spreading
→ macro productivity gains are still narrow
→ policy and investment should track diffusion quality
Workplace Evidence: What AI Actually Changed
The QJE version of Generative AI at Work studies 5,172 customer-support agents at a Fortune 500 firm using a GPT-3-based conversational assistant. The tool offered real-time response suggestions that agents could accept, edit, or ignore.
The headline result is a roughly 15% increase in issues resolved per hour. But the deeper result is distributional. Lower-skilled and newer workers benefited most, while experienced high performers saw smaller speed gains and some quality trade-offs.
This matters because AI did not merely make the best workers even better. It compressed skill gaps.
| Finding | Economic meaning |
|---|---|
| Larger gains for less experienced workers | Faster onboarding and lower training cost |
| Text style shifted toward high-skill agents | Tacit knowledge became more transferable |
| Customer sentiment improved | Quality and work experience also changed |
| Stronger effects for moderately rare issues | AI helps when the firm has patterns but humans lack repeated exposure |
This is not full automation. It is workflow augmentation. The first-order effect is not “remove the human”; it is “raise the average worker’s output and quality.”
Adoption: AI Is Already Inside the Workplace
The Fed’s FEDS Notes reconciles three different survey lenses. The numbers vary because they use different units, but together they show that AI has moved beyond experimentation.
| Estimate | Unit | Interpretation |
|---|---|---|
| 18% | Firm count | Share of firms reporting AI adoption at year-end 2025 |
| 41% | Individuals | Workers reporting work-related GenAI usage |
| 78% | Employment-weighted firms | Workers employed at AI-adopting firms |
| 54% | Employment-weighted LLM firms | Workers employed at firms using LLMs |
The high employment-weighted numbers matter. Large firms employ many workers, so AI adoption by large organizations can reach a large share of the workforce even if the firm-count adoption rate looks modest.
The industry pattern also matters. AI adoption is strongest in finance and professional services, the high-value cognitive work sectors where research, analysis, compliance, coding, legal drafting, accounting, and client communication are core workflows.
Macro Productivity: Promising but Not Broad-Based Yet
The Kansas City Fed provides the necessary caution. U.S. labor productivity has strengthened since late 2022 and moved above its pre-pandemic trend. That period overlaps with the commercial emergence of widely used generative AI tools.
But the pickup is not broad-based. A relatively small set of industries accounts for much of the improvement. Higher AI adoption is associated with faster productivity growth across industries, but it explains little of the aggregate productivity shift so far.
That is not a bearish conclusion. It is an early-diffusion conclusion.
Task-level gains
→ firm workflow redesign
→ industry-level adoption
→ macro productivity data
The economy is still moving through the middle of that chain.
Policy Lesson: Do Not Wait for Aggregates
The San Francisco Fed frames the AI moment through the historical lens of the 1990s productivity acceleration. The lesson is not that AI will repeat the 1990s exactly. The lesson is that aggregate data often lag transformative business change.
For monetary policy, AI is complicated because it can create short-term demand pressure and long-term supply benefits at the same time.
| Channel | Short-term effect | Long-term possibility |
|---|---|---|
| Data-center capex | Raises demand for power, equipment, land, financing | Builds productive capacity |
| AI tools | Implementation costs and disruption | Lower unit labor cost |
| Labor market | Task displacement and job redesign | Higher output per worker |
| Inflation | Infrastructure bottlenecks | Disinflation from productivity |
This is why the Fed cannot simply wait for aggregate productivity to settle the debate. It needs firm-level and industry-level signals.
Investment Read-Through
For investors, the key implication is that AI capex can still be justified, but not by model benchmarks alone. It needs workflow productivity.
Positive evidence:
- real task-level productivity gains
- rising work-related adoption
- early penetration into high-value cognitive sectors
- signs of tacit knowledge transfer and faster onboarding
Remaining risks:
- usage does not equal value creation
- workflow redesign takes time
- aggregate productivity gains remain narrow
- capex may outrun proven ROI
- quality, compliance, and liability issues can slow adoption
For Korea, this supports the long-term logic behind AI infrastructure: HBM, server DRAM, eSSD, networking, power equipment, data-center construction, AI cloud, and enterprise workflow software. But it also raises the bar. The market should reward companies that turn AI into repeatable productivity workflows, not companies that merely attach AI labels to old businesses.
Final View
AI productivity is real at the task level. Adoption is high enough to matter. But the macro productivity revolution is not yet fully proven.
That is exactly why the next phase of the AI cycle is about diffusion, not demo performance. The question is no longer whether AI can answer questions. The question is whether it changes workflows deeply enough to show up in margins, output per worker, and eventually the national productivity data.