Anthropic's AI: Specific Use Cases Dominate

Anthropic's analysis of the use of large language models (LLMs) indicates a concentration on a limited number of tasks. Approximately a quarter of consumer interactions and a third of enterprise API traffic focus on the ten most frequent activities. Code creation and modification are among Claude's primary uses.

This trend suggests that the model's value lies primarily in these areas, without significant expansion into other uses. Therefore, targeted AI implementations, focused on tasks where language models have proven effective, may prove more effective than generalized implementations.

Augment, Don't Just Automate

On consumer platforms, collaborative use, where users iterate queries to the AI, is more common than full automation of workflows. Enterprise API usage shows the opposite, with companies seeking to save by automating tasks. However, the quality of results decreases with the complexity of the task and with longer processing times.

Automation is most effective for simple and well-defined tasks that require few logical steps and quick responses. For longer tasks, users must iterate and correct outputs. Breaking down complex tasks into more manageable phases improves results.

Most queries to LLMs come from professionals, although in less developed countries Claude is used more often in academia. Some professionals, such as travel agents, can delegate complex planning tasks to AI, while maintaining transactional activities. Conversely, property managers can automate routine administrative tasks, while maintaining activities that require greater judgment.

Productivity: Gains are Reduced by Reliability

Estimates of a 1.8% increase in labor productivity thanks to AI should be reduced to 1-1.2%, considering the extra costs and labor required for validation, error handling, and rework. A 1% increase is still significant, but business decision-makers must take these factors into account.

Potential gains also depend on whether AI complements or replaces human work. Replacement depends on the complexity of the task. An almost perfect correlation was found between the sophistication of prompts and the success of the results: the use of AI determines what it provides.

Conclusions for Leaders

  • AI implementation generates value more quickly in specific and well-defined areas.
  • Complementary systems (AI + human) outperform full automation for complex jobs.
  • Reliability and extra work reduce expected productivity gains.
  • Changes in the composition of the workforce depend on the combination of tasks and their complexity, not specific roles.