Cohere mapped almost 700,000 agent tools. Only 2.6 percent can complete a work task from start to finish

Cohere's research provides a useful counterweight to broad claims about rapid automation of work. The team collected 696,291 tools from 123,069 public server records based on the MCP protocol and attempted to match them to specific work activities in the U.S. O*NET database. It counted as genuine automation only a tool that performs the activity from beginning to end, not a system that merely gives a person information or performs one partial step. Under this strict definition, about 2.6 percent of tools were found to perform an existing work task independently, roughly one in forty. The matches are also heavily concentrated in digital occupations. In 419 of the 923 occupations studied, researchers found no public agent-tool activity. Cohere therefore warns that the future of work will not be determined by AI capability alone. The availability of usable tools and their connection to real processes are equally important. The figure of 2.6 percent cannot, however, be interpreted as the share of work that can be automated. The study examines offerings in public directories, not all internal enterprise systems. Cohere explicitly acknowledges that non-public servers may contain more automation of administrative and operational activities. The authors therefore treat the result more as a lower bound for visible supply. In addition, a large number of similar tools may cover a single task and says nothing about their actual use. For company leadership, the most useful distinction is between 'can help' and 'can take over the process.' An assistant that creates a draft can save minutes, but still leaves a person coordinating the whole procedure. Agentic automation begins where a system takes over several consecutive steps, works with company systems and has a clear completion condition. That is where a larger productivity benefit emerges, but also a greater need for access rights, control and accountability. For public policy, the study cautions against simple forecasts of occupations disappearing based only on model capability. The transition from technical capability to economically usable automation depends on integration, price, responsibility, data and organizations' willingness to change processes. Employers are therefore better served by mapping individual activities and their dependencies than by labeling entire occupations 'automatable.' This also makes it easier to see where investment should go into retraining and where it should go into technical integration. The study also suggests why corporate experience with AI differs so much across departments. Where work is digital from the outset, standardized and well documented, agent tools emerge faster. In physical, relationship-based or heavily regulated occupations, integration is slower. The difference may therefore say less about the model's 'intelligence' than about how easily a workflow can be translated into interfaces, data and a machine-verifiable result. KEY TERMS MCP: A protocol that allows models to communicate with external tools and data sources in a standardized way. O*NET: A U.S. database of occupations, work activities, skills and requirements. End-to-end automation: Taking over an entire defined activity rather than merely helping with one step.

Cohere mapped almost 700,000 agent tools. Only 2.6 percent can complete a work task from start to finish

The full article is available to users registered at Hard Skills