Article reader Listen + reading controls
Article reader
Preparing the reader…
Reading settings
AI adoption begins with tasks, not job titles¶
Conversations about artificial intelligence and work often begin at the wrong level. They ask which jobs will disappear, then argue over forecasts that are too coarse to guide an actual organization.
Anthropic's first Economic Index, published February 10, analyzes how people use Claude across occupational tasks. The report finds usage concentrated in particular kinds of work and distinguishes between automation, where the model performs a task, and augmentation, where people and the model work together.
Jobs are bundles¶
A job title hides a bundle of activities. A program manager may interpret requirements, negotiate resources, build a schedule, write a briefing, resolve conflict, coach staff, and notice risks that no dashboard captures. A software engineer may design interfaces, trace a production failure, write tests, review code, and explain tradeoffs to a mission owner.
Artificial intelligence (AI) will not affect each activity equally. Some tasks are easy to specify and verify. Others depend on local history, physical context, trust, authority, or tacit knowledge. Treating the whole job as automatable or immune obscures where change will actually occur.
Research on task-based technological change has long made this distinction. Autor's analysis of tasks and labor argues that automation usually substitutes for some human activities while complementing others. The organizational outcome depends on how work is recomposed.
Productivity can move the work¶
Suppose AI reduces the time required to produce a first draft of code, analysis, or documentation. The saved time does not disappear into an abstract productivity measure. It changes queues and handoffs.
Review may become the bottleneck. Domain experts may receive more material to validate. Junior staff may produce more while encountering fewer opportunities to struggle with foundational problems. Managers may have to decide which work deserves human attention when the cost of generating options approaches zero.
This is why seat counts and usage rates are weak measures of adoption. A tool can see heavy use while the surrounding system accumulates review debt, duplicated work, uncertain provenance, and fragile skill development.
Map the work before redesigning it¶
An organization can use a task-level approach without launching a year-long study. For a priority workflow:
- Observe how the work is actually done, including exceptions and informal handoffs.
- Separate information transformation from judgment, commitment, and relationship work.
- Identify where quality can be checked cheaply and where failure is difficult to detect.
- Decide whether AI should draft, advise, challenge, execute, or remain outside the task.
- Redesign review, training, and performance measures around the new allocation.
- Track whether cycle time, quality, learning, and workload improve together.
The distinction between automation and augmentation is helpful, but neither is automatically good. Augmentation can bury people under suggestions. Automation can remove low-value work or eliminate the practice through which expertise develops.
The right question is not “What percentage of this job can AI do?” It is “What arrangement of people, AI, process, and evidence will produce better outcomes while preserving the capabilities the organization still needs?”
That is a design question. The task data gives us a better place to begin.
Sources and research trail¶
- Anthropic, “The Anthropic Economic Index” (February 10, 2025).
- Autor, “Why Are There Still So Many Jobs? The History and Future of Workplace Automation” (2015).
- Parker and Grote, “Automation, Algorithms, and Beyond: Why Work Design Matters More Than Ever” (2020).
- Faraj, Pachidi, and Sayegh, “Working and Organizing in the Age of the Learning Algorithm” (2018).