← Back to perspectives
/Intellicon FDE Team

Anthropic Finds Robots Are Cheaper Than Human Workers for Only 0.3% of Tasks

The analysis covers 900 occupations and 19,000 tasks, factoring in maintenance, electricity, and human oversight. When evaluating an AI Agent, businesses should likewise track human review, rework, and exception handling.

AIEgentWrX人機協作導入
Anthropic Finds Robots Are Cheaper Than Human Workers for Only 0.3% of Tasks

A sales representative tests an AI Agent with a request for quotation. The AI Agent retrieves historical pricing, payment terms, and delivery schedules, then drafts the quotation. At the meeting, the manager’s next question is inevitable: Now that it works, should we make it available to more employees—or even integrate it into formal workflows?

A successful demonstration proves only that the technology is viable. The business must still determine how much each run costs in model usage, how much time employees spend reviewing the output, how often errors require rework, and who takes over when exceptions arise. Without these records, managers cannot reliably estimate the costs or risks of scaling adoption.

If It Works, Why Might It Still Not Be Cost-Effective?

In its coverage of Anthropic’s report on robots’ ability to perform work, TechOrange noted that the study used the U.S. Department of Labor’s O*NET occupational database, covering 900 occupations and approximately 19,000 detailed tasks. Claude then helped analyze and score each task. The results showed that current robotics technology can perform 74% of physical tasks in the United States, representing about 34% of total work hours. However, once equipment, maintenance, electricity, and human oversight are included, only 0.3% of tasks are cheaper to perform with robots than with human workers.

The study is based on U.S. occupations, wages, and deployment costs, so its percentages cannot be applied directly to Taiwanese businesses. Its evaluation method, however, is highly relevant to AI Agent projects. During a demonstration, teams usually focus only on whether the AI Agent can retrieve information, generate documents, and invoke tools. Once it enters actual use, the company must also bear the costs of model usage, system integration, access management, human review, error-related rework, and exception handling. If the workflow involves sending external emails, approving prices, or modifying data, the potential impact of errors on customers and internal operations must also be included in the risk assessment.

Before approving the next phase, managers should therefore require the project team to provide empirical records, including the number of tokens used per run, the time spent on human review, the circumstances that required rework, and the exceptions that still had to be handled by employees. Before scaling adoption, the organization must have actual records of the full costs, risks, and human effort involved—not merely an impressive demonstration.

Model Charges Are Only Part of the Cost

To evaluate the cost of an AI Agent, begin with a clearly defined task that is easy to validate, such as organizing sales inquiries, consolidating procurement price-comparison data, or checking report fields against quality-control requirements. The team should define the input data, expected output, and acceptance criteria in advance, then have the same group of users follow a standardized process during the pilot. Only then can the differences between individual runs be compared meaningfully.

Model charges are a direct cost, but the time employees spend preparing data, reviewing content, returning work for revision, and handling exceptions also affects whether the project is ultimately cost-effective. If a sales representative still needs to verify every price, term, and delivery date after the AI Agent drafts a quotation, the project team must record that review time. If historical data is stored in inconsistent formats and IT or sales staff must first spend time organizing it, those labor hours must also be included in the implementation cost. Procedures and responsible personnel should also be defined in advance for situations such as emails sent to the wrong recipient, outdated terms being cited, or unauthorized data being accessed. These rules should not be added only after the system goes live.

During the pilot, teams should consistently record five categories of data: model and tool usage, time spent on human preparation and review, the number of rework cycles, exception-handling details, and errors that affect customers or internal operations. They can then compare the full weekly or monthly cost against the labor requirements and potential risks of the existing process to decide whether to continue the pilot, revise the workflow, or scale up adoption.

EgentWrX provides a four-tier budgeting mechanism covering tenants, departments, members, and AI Agent, with the alert threshold set to 80% by default. If any tier exhausts its budget, the system will block subsequent use only when the administrator has selected the “Block” option. Project teams should set budget caps for both the pilot department and the AI Agent from the outset, rather than waiting for charges to rise before investigating where the money was spent.

Which Steps Require Human Approval?

The study covered by TechOrange includes human oversight in the deployment cost of robots. It also notes that even when automated welding equipment can perform the primary operation, workers are still needed on-site to move metal, climb to position components, and grind surfaces afterward. AI Agent workflows present a similar situation: data organization and content drafting can be assigned to an AI Agent, but whenever money, accountability, or external impact is involved, the company must still designate a person to make the final decision.

Project teams can divide the workflow into three categories. The first consists of steps that an AI Agent can perform according to explicit rules, such as organizing inquiry fields or comparing them against existing terms. The second consists of steps that require human approval after the AI Agent completes its work, such as when a discount exceeds the department’s limit, different data sources conflict, or a quotation is about to be sent to a customer. The third should remain entirely under employee control, including situations where the terms are ambiguous, customer relationships are involved, or an exception falls outside existing company policies.

Human approval should not be reduced to a signature field at the end of the workflow. The company must define in advance the conditions under which a case may proceed, who is authorized to approve it, and what should happen if no action is taken by the deadline. During the pilot, teams should also record what approvers changed, why work was returned, how often approved actions still had to be reversed, and how much time was spent on remediation. The workflow should then be adjusted based on these empirical results.

EgentWrX allows highly repetitive work to be configured as tasks, with multiple tasks linked into a workflow. If human approval is required before a handoff, the workflow pauses until the designated person authorizes it to proceed. A business can begin with a sales quotation or procurement price-comparison workflow, clearly mapping which steps can be performed by an AI Agent, which require human approval, and which remain the responsibility of employees. It can then establish budget caps and a defined pilot period. At the review meeting, managers should examine the full cost, exception records, and approval outcomes—not merely whether the AI Agent can produce a document that appears complete.

FAQ

If an AI Agent can already complete the work, is a proof of concept still necessary?

Yes. In addition to confirming whether the AI Agent can produce the desired result, a proof of concept (PoC) should record model usage, human review time, the number of rework cycles, and how exceptions are handled. A business can begin with a clearly scoped task that is easy to validate, run the actual workflow using production data, and then decide whether to scale up adoption.

What should be included in the cost of an AI Agent?

At a minimum, the calculation should include model and tool usage, system integration, data preparation, access management, human review, rework, and exception handling. If the workflow involves sending external emails, approving prices, or modifying data, it should also include the labor required to reverse actions and remedy errors.

Which steps should require human approval?

The workflow should pause for review by a designated person when an amount exceeds the department’s limit, data sources conflict, terms are unclear, or content is about to be sent externally. The business must also specify who is authorized to approve the action, who is responsible for revisions after a rejection, and what should happen if the matter remains unresolved past the deadline.

Can budget alerts alone prevent cost overruns?

No. In addition to configuring alerts, administrators must choose to block subsequent use when the budget is exhausted and set separate limits for tenants, departments, members, and AI Agent. During the pilot, the organization should also review actual usage regularly to confirm that the budget caps align with operational requirements and the costs the company is prepared to accept.

References

30 minutes to map out which work to hand to AI first

Want every employee to have their own AI teammate?

A consultant will be in touch shortly.

ISO/IEC 27001:2022 Certified ISMS