How much can generative AI improve workplace productivity?
Research found 40% less writing time and 15% more support issues resolved per hour with generative AI. Compare gains by task and experience, measure your own trial and review Korea’s AI Basic Act.
What is the main conclusion on generative AI adoption?
Controlled experiments have found clear productivity gains from generative AI. Professional writing tasks took 40% less time and output quality improved by 18%, while customer support agents resolved an average of 15% more issues per hour. However, the effects varied substantially by the work involved and the user's level of experience. The sensible first step is therefore to measure a small trial in your own company's tasks.
How large were the productivity gains in the research?
To assess the effects without exaggerating them, look for studies with comparison groups, such as randomized assignment or phased adoption. The designs and findings of two prominent studies are summarized below.
Two studies measuring generative AI's productivity effects
| Study | Participants and design | Main findings |
|---|---|---|
| Noy and Zhang (2023) | 453 college-educated professionals completed occupation-specific writing tasks; half were randomly given access to ChatGPT | Task completion time fell by 40% and output quality improved by 18% |
| Brynjolfsson, Li and Raymond (2025) | An AI conversational assistant was introduced in phases to 5,172 customer support agents | Issues resolved per hour increased by 15% on average, with substantial differences across workers |
Both studies measured outcomes in specified tasks and jobs. Their results should not be interpreted as equivalent increases in the productivity of an entire company. In the writing experiment, participants who tried ChatGPT were twice as likely to report using it in their actual work two weeks later, and 1.6 times as likely two months later. A small practical trial may therefore help wider adoption get started.
Why does the same AI tool affect people differently?
In the customer support study, less experienced and lower-skilled agents improved both their speed and quality. More experienced and higher-skilled agents saw a small increase in speed but a small decline in quality. The writing experiment also reduced performance differences between participants. The support study additionally reported larger effects on unusual problems that agents encountered infrequently, together with improvements in the experience of work: customers were more polite and less likely to request a manager.
These findings have implications for adoption strategy. If an AI tool can transfer experienced workers' know-how to beginners, onboarding and standardizing work may be good places to try it first. Experienced workers, by contrast, may benefit from choosing the scope of use themselves so that the tool does not interfere with their judgment.
Can research findings be transferred directly to our company?
The conditions of an experiment are not identical to those in your company. Different tasks, reviewers and customer expectations can produce different effects. When referring to another company's results, the Korea Research Institute of Business & Branding makes it a principle to explain differences in industry, size, timing and market conditions. This principle is discussed further in how to use other companies' success stories.
Which measures should we use to assess AI adoption?
Measuring processing time alone can miss a decline in quality or an increase in review work. Define your performance indicators before adoption and compare the periods before and after using the same criteria.
Examples of indicators for measuring AI adoption
| Indicator | What it measures | Risk to examine alongside it |
|---|---|---|
| Processing time | Time required to finish the same task | Whether the time saved is offset by longer review |
| Output quality | Scores assessed against predefined criteria | The difference between apparent polish and factual accuracy |
| Throughput | Tasks completed in the same amount of time | Whether only the number of easy tasks has increased |
| Rework rate | The share of outputs requiring correction | Whether errors are being passed to later stages |
| Employee experience | Perceived workload and learning | Whether experienced workers' judgment is weakening |
Starting with a small trial in one team and expanding after checking the indicators follows the same principle as testing new business hypotheses.
What organizational preparation is needed before adoption?
Introducing a tool alone does not improve productivity. Decide which tasks AI will perform and which people will perform, who will review outputs and how, and how evaluation criteria should change. The Korea Research Institute of Business & Branding strengthened its research on AI use and organizational productivity in April 2026, examining the economics of technology adoption, changes in work and conditions for implementation alongside organizational operations. Teaching people how to check the facts in outputs, as well as how to use the tool, helps prevent the review burden from falling on one person. For ways to review work processes and people systems together, see organizational diagnosis and operational improvement. For connections with other areas, see business research and advisory services.
What does the AI Basic Act that took effect in 2026 require of businesses?
The Framework Act on the Development of Artificial Intelligence and Establishment of a Foundation for Trust took effect on January 22, 2026. The Act defines AI business operators as businesses that develop and supply AI and businesses that use it to provide products or services. Providers of products or services using high-impact AI or generative AI must notify users in advance of that use, and outputs from generative AI must be identified as generated content.
AI used for decisions that significantly affect individuals' rights and obligations, such as hiring or loan screening, falls within the area of high-impact AI. If you plan to move beyond internal work and offer AI-based customer services or use AI in recruitment assessments, review whether those provisions apply in advance.
Frequently asked questions about generative AI adoption
Which tasks should we introduce generative AI into first?
It is easier to measure effects when starting with repetitive tasks whose outputs people can readily review. Drafting documents or customer support responses are common starting points.
Does AI reduce the role of experienced employees?
The studies found smaller speed gains among experienced workers, but their role in reviewing outputs and setting standards may grow. Responsibilities need to be reassigned as part of adoption.
How long does it take to establish whether adoption is working?
That depends on the work cycle. Measure a baseline before adoption, then compare after the same tasks have been repeated sufficiently to distinguish temporary changes from real effects.
How should we manage factual errors in generative AI outputs?
The basic approach is to include human verification in the workflow instead of releasing outputs unchanged. A checklist of error-prone elements such as figures, proper names and quotations can reduce review time.
Should we check the AI Basic Act even if AI is used only internally?
The Act's transparency duties apply to businesses providing AI-based products or services. Recheck applicability if use expands into customer services or recruitment assessment.
References
- Science, “Experimental evidence on the productivity effects of generative artificial intelligence” (Noy and Zhang), 2023. https://doi.org/10.1126/science.adh2586
- The Quarterly Journal of Economics, “Generative AI at Work” (Brynjolfsson, Li and Raymond), 2025. https://doi.org/10.1093/qje/qjae044
- National Law Information Center, Framework Act on the Development of Artificial Intelligence and Establishment of a Foundation for Trust, effective January 22, 2026. https://www.law.go.kr/법령/인공지능발전과신뢰기반조성등에관한기본법