# Evidence That AI Agents Are Replacing Knowledge Workers Remains Thin

> AI is being woven into knowledge work and operations, yet large-scale layoffs, net job losses, and measured productivity gains remain unconfirmed.

Published: 2026-09-23T18:33:58.594Z
Updated: 2026-09-23T18:33:58.594Z
URL: /en/article/ai-agents-work-replace-2026

# The Future of AI Agents and Knowledge Work: Redesigning Jobs Comes Before Replacing Workers

## Background

The debate surrounding artificial intelligence has long centered on the question, “Will AI take people’s jobs?” However, taken together, the available materials point to a more immediate and clearly observable change: rather than large-scale workforce replacement, some tasks within knowledge work are being broken down into units that can be automated and then redesigned within corporate operations. In other words, the key signal is not the replacement of entire workforces, but the automation and restructuring of individual tasks.

Boston Consulting Group (BCG) describes applied AI as a means of building organizational capabilities and scaling operations. In this context, AI is presented not merely as an office productivity tool, but as a technology connected to a wide range of functions, including schedule optimization, the execution and launch of innovative ideas, supply chain management, and operational efficiency. This suggests that companies are not viewing AI solely as a mechanism for eliminating specific jobs. Instead, they are using it to break existing workflows into smaller components and automate parts of decision-making and execution.

That said, BCG’s reported figures, including more than $4.5 billion in profit-and-loss improvements and various productivity and efficiency metrics, require careful interpretation. According to the excerpts provided, these figures were presented as company-reported examples. Their independent verification status, calculation methodology, and comparison periods are not clear. It is therefore difficult to generalize them across entire industries or conclude that AI adoption will automatically produce results of a similar scale.

Signals from model providers also point to products aimed at knowledge work. Anthropic introduces Fable 5.1 and Mythos 5.1 as models for coding and knowledge work, while describing Sonnet 5 as a more agentic model. This suggests that tasks such as research, writing, and coding are regarded as major application areas for AI agent products. However, the release of a particular model does not mean that these tasks can be completed in real-world settings without human involvement. The performance, speed, and cost claims presented by Anthropic also have not been independently verified based on the materials provided.

## Key Issues

The first key issue is whether **AI agents replace entire occupations or automate only parts of people’s work**. The conclusion most directly supported by the current materials is closer to the latter. BCG’s examples focus on improving operations across business functions, while Anthropic’s product descriptions demonstrate the possibility that models can perform certain tasks in knowledge work, such as coding, research, and writing. However, the materials do not provide direct evidence of which occupations have disappeared entirely, how much net employment has declined since the adoption of AI, or the scale of AI-related layoffs.

Work is not a single, indivisible unit. It consists of multiple stages, including drafting documents, searching for information, organizing data, coordinating schedules, reviewing drafts, entering information into external systems, and giving final approval. AI agents can automate repetitive and highly rule-based steps, but taking responsibility for an entire workflow is a separate matter. Judging the quality of external data, making decisions based on an organization’s goals and risk tolerance, and assuming responsibility when errors occur are difficult to evaluate solely in terms of the level of automation.

The second key issue is **the outcome of productivity gains**. Even if a company reduces processing time by adopting AI, the result does not necessarily lead to workforce reductions. The savings may instead be used to retain employees, handle a larger volume of work, expand services, increase quality-control staff, or develop new products. The available materials do not provide tracking data that would indicate which of these outcomes is most common.

The third key issue is **the hidden cost of AI adoption**. The list of topics from Yale Insights includes AI content crawling, junior employee training, community benefits from data centers, and research and commentary related to agentic AI. This shows that AI is expanding beyond a simple software function into issues involving information collection and knowledge distribution, talent development, and physical computing infrastructure. Business Insider likewise covers the impact of data centers on businesses, the environment, local communities, and employment as a major area of reporting.

However, the list alone cannot establish whether the effects of data centers or AI content crawling are positive or negative, or determine their scale. Identifying the direction of an issue is different from measuring its impact. As AI agents collect more information and execute more services, discussions about computing resources, data management, workforce training, and the burden placed on local communities may also intensify.

## A Multidimensional Analysis

### 1. The Corporate Operations Perspective: Processes Change, Not Just Jobs

An important point in BCG’s framing is that AI is treated less as a technology that replaces specific corporate roles all at once and more as a capability embedded across operations. The processes involved in schedule optimization, supply chain management, operations, and the launch of innovative ideas span multiple departments and interconnected tasks. When AI automates parts of these processes, the result may be more than a simple change in headcount. It can also alter the sequence of work and the structure of decision-making.

For example, when repetitive information gathering or draft writing is automated, employees may spend more time on review, exception handling, and prioritization. At the same time, new tasks may emerge to monitor automated outputs and correct errors. As a result, it is difficult to explain the full extent of organizational change by asking only, “How many people did AI replace?” Workload, the substance of roles, the scope of responsibility, and rework rates must also be examined.

### 2. The Model Provider Perspective: Agentic Products Expand the Scope of Execution

Anthropic’s introduction of models for coding and knowledge work, along with more agentic models, is a product signal that the generative AI competition is moving from producing answers to performing real-world tasks. Agentic AI is designed not merely to respond to questions, but to connect a series of activities, including research, writing, and coding.

However, there are several stages between the existence of a product and actual labor replacement. Whether a model can perform a specific task, connect reliably to a company’s data and tools, how often humans must correct its errors, and whether security and access controls can be maintained are separate issues that require validation. The available materials do not provide data on agents’ actual task-completion rates or long-term performance comparisons in controlled environments. Product launches should therefore not be interpreted automatically as evidence of labor-market disruption.

### 3. The Information Ecosystem Perspective: The Foundations of Knowledge Work Are Changing

The topics highlighted by Yale Insights suggest that AI’s impact extends beyond the office itself. AI content crawling is related to the structure of the internet from which models learn or collect information, while junior employee training is connected to the role that entry-level work plays in developing talent. If some entry-level tasks become automated, organizations may need to redesign how new employees gain experience.

This is not simply a productivity issue. Knowledge workers often learn an organization’s context and standards of judgment by performing repetitive tasks. If that process becomes automated, efficiency may improve in the short term, but the pathways through which people receive training and develop expertise may change over the long term. However, the available materials do not establish the actual consequences of these changes, so it is important to distinguish between possibilities and verified facts.

### 4. The National Security Perspective: Efficiency and Control Risks Grow Together

The Belfer Center’s focus on AI in the context of national capabilities, international security, U.S.-China competition, and AI-augmented government operations shows that the spread of AI agents cannot be viewed solely through the lens of corporate productivity. When agents are connected to browsers, internal company data, and external services, they can perform more tasks. At the same time, the risks of privilege abuse, information leakage, and erroneous execution can also increase.

From this perspective, the central question about AI is not simply, “How many tasks can be automated?” It is also, “With what permissions, accessing which data, and with whom responsible for the outcome?” Greater efficiency can turn into amplified risk when oversight systems are weak. AI agent adoption is therefore not merely a technology procurement decision. It is also a governance challenge involving permission design and the allocation of responsibility.

## Outlook

The most cautious yet meaningful conclusion supported by the current materials is that **some forms of knowledge work are moving into products and infrastructure designed for automation**. In corporate operations, AI is being integrated into functions such as scheduling, supply chain management, innovation, and operations. Model providers are also releasing agentic products aimed at knowledge work, including coding, research, and writing. At the same time, the topics covered by Yale Insights, the Belfer Center, and Business Insider show that AI’s impact is expanding into labor, the information ecosystem, data centers, local communities, governance, and international security.

However, this is not enough to claim that “AI agents have replaced knowledge workers on a large scale.” There is no direct evidence regarding net employment declines, the scale of layoffs, controlled productivity comparisons, or actual end-to-end task completion rates. BCG’s performance examples and Anthropic’s product descriptions are important signals, but each has limitations: the former consists of company-reported business cases, while the latter represents product claims from a model provider.

Looking ahead, three verification challenges will be especially important. First, organizations need real-world comparisons that disclose task-level processing times, error rates, rework rates, and labor costs before and after AI agent adoption. Measurement should focus not simply on usage rates or the number of adopting companies, but on actual work outcomes. Second, researchers must track whether productivity gains result in workforce reductions, employment retention, or increased workloads. Third, it is necessary to measure how often privilege abuse and declines in information quality occur when agents are connected to browsers, internal company data, and external services.

Ultimately, the future of AI agents will not be determined by technical performance alone. It will also depend on how companies reallocate the time saved through automation, how organizations train entry-level employees, who bears the costs of data centers and information infrastructure, and who is held responsible for errors and harm. The appropriate approach today is neither optimism nor fear, but **carefully observing changes at the task level while separating them from employment outcomes**. AI is already becoming integrated into knowledge work, but there is still insufficient evidence to conclude that this integration necessarily means large-scale job replacement.

## Claims

- AI는 기업 운영과 생산성 개선, 일정 최적화, 공급망, 혁신 업무에 적용되고 있다. (확인됨)
- BCG는 45억 달러 이상의 손익 개선과 일부 사례의 생산성·효율성 개선을 제시하지만, 독립 검증이나 산출 방법은 아직 확인된 바 없다. (부분 확인·검증 미완료)
- Anthropic은 코딩과 지식노동을 겨냥한 모델 및 더 에이전트적인 모델을 제품 발표에서 소개한다. (확인됨)
- Anthropic이 제시한 모델의 성능·속도·비용 우위는 제공 자료에서 독립적으로 검증되지 않았다. (검증 미완료)
- Belfer Center는 AI를 국가운영, 국제안보, 신흥기술, 미·중 경쟁의 맥락에서 다룬다. (확인됨)
- Business Insider는 데이터센터가 기업·환경·지역사회·일자리에 미치는 영향을 보도 영역으로 제시한다. (확인됨)
- AI 에이전트가 지식노동자를 대규모로 대체했거나 순고용을 줄였다는 점은 아직 확인된 바 없다. (확인됨)

## Forecasts

- 75% — 에이전트 도입은 먼저 전체 직무가 아니라 반복적이고 측정 가능한 업무 단위에서 확대될 것이다 (12개월). Signal: 기업 발표와 독립 사례에서 코딩, 일정관리, 리서치, 문서작성, 고객지원 등 특정 업무별 처리시간·오류율·검수시간의 전후 비교가 공개되는지 확인한다.
- 65% — 기업의 AI 도입 공시보다 실제 순고용 감소의 증거가 늦게 나타날 것이다 (18개월). Signal: AI 도입 기업의 동일 직군 채용공고, 주니어 채용, 계약직 비중, 해고·재배치 수치를 도입 전후로 비교한다.
- 70% — 에이전트 성능 경쟁은 모델 자체보다 데이터·권한·검수 인프라 경쟁으로 이동할 것이다 (24개월). Signal: 기업이 모델명보다 사내 데이터 연결, 브라우저·도구 권한, 감사 로그, 사람 승인 절차, 오류 복구율을 핵심 구매 기준으로 공개하는지 추적한다.
- 60% — 데이터센터와 AI 인프라의 지역사회·에너지 비용이 도입 속도를 제한하는 반대 신호가 될 것이다 (24개월). Signal: 데이터센터 인허가 지연, 전력·용수 비용, 지역사회 반대, 환경 영향 검토가 기업의 에이전트 확장 계획을 변경하는 사례가 늘어나는지 확인한다.

## Sources

- [Strategic Management Consulting | Boston Consulting Group](https://www.bcg.com/) — bcg.com, 2026-09-23
- [Yale Insights | Ideas from the Yale School of Management](https://insights.som.yale.edu/) — insights.som.yale.edu, 2026-09-23
- [The Belfer Center for Science and International Affairs | Harvard Kennedy School](https://www.belfercenter.org/) — belfercenter.org, 2026-09-23
- [Home | Hub](https://hub.jhu.edu/) — hub.jhu.edu, 2026-09-23
- [AIMultiple: AI Use cases & Tools to Grow Your Business](https://aimultiple.com/) — aimultiple.com, 2026-09-23
- [Carnegie Endowment for International Peace | Carnegie Endowment for International Peace](https://carnegieendowment.org/) — carnegieendowment.org, 2026-09-23
- [Platformer](https://www.platformer.news/) — platformer.news, 2026-09-23
- [Fortune - Fortune 500 Daily & Breaking Business News](https://fortune.com/) — fortune.com, 2026-09-23
- [Home \ Anthropic](https://www.anthropic.com/) — anthropic.com, 2026-09-23
- [DemandSage - Data Reporting & SaaS](https://www.demandsage.com/) — demandsage.com, 2026-09-23
- [Tech Insider](https://tech-insider.org/) — tech-insider.org, 2026-09-23
- [The Economy — Institutional Intelligence for the AI Economy](https://economy.ac/) — economy.ac, 2026-09-23
- [EPIC for America | Economic Policy Innovation Center](https://epicforamerica.org/) — epicforamerica.org, 2026-09-23
- [Business Insider - Latest News in Tech, Markets, Economy & Innovation](https://www.businessinsider.com/) — businessinsider.com, 2026-09-23