Cohere Labs Maps the Agentic Task Ecosystem — Only 2.6% of 696,000 AI Tools Fully Automate Work Tasks

Key Takeaways

→ Only 2.6% of 696,000 published AI tools can fully carry out a recognized work task end-to-end

→ 419 of 923 US occupations (nearly half) have no agentic tools built for them at all

→ What gets built tracks expert feasibility judgments, not what workers want automated

→ Healthcare and computing see automation of specialized tasks; legal, production, and sales see only routine edges automated

→ Only ~3% of tool categories represent genuinely new work — mostly managing AI agents themselves

Cohere Labs has released the Agentic Task Ecosystem (ATE) dataset — roughly 696,000 published AI tools across 123,000 MCP servers, the largest open dataset of its kind — and its findings complicate the narrative that AI agents are rapidly swallowing occupations.

Key Findings

Only 2.6% of tools clear the bar — under a strict test asking whether a tool fully carries out an occupational task (rather than merely informing a person doing it), about one tool in forty qualifies
419 of 923 U.S. occupations have no agentic tools at all — for anyone asking which jobs agentic AI is “coming for,” the honest supply-side answer is that for nearly half of them, nobody is building anything yet
Coverage is uneven — tools cluster in software/IT work and information-structuring roles; graphic designers have tools matching 11 of their 15 software-performable tasks
What gets built follows feasibility, not preference — expert judgments of technical feasibility predict which occupations receive tools; workers’ own preferences about what they want automated predict nothing

Where Automation Lands Inside Jobs

The study’s most nuanced finding: it’s not how much of an occupation is automated, but which parts.

Healthcare and computing — tools reach toward the specialized end, leaving humans the routine remainder (“expertise-lowering automation”)
Legal, production, and sales — tools stay at the routine edges, leaving the specialized core with people (“expertise-raising automation”)

Specialized work resists automation when it’s physical or interpersonal — and gives way when it’s already conducted through software. Clinical Data Managers (143 tools) and Biostatisticians (82 tools) are the clearest cases of specialized information work attracting heavy automation.

The Other 98%

Grouping unmatched tools by similarity, the researchers found:

Mostly existing work at the wrong grain — “subatomic” tools that do pieces of tasks, or “composite” tools bundling several tasks
Agent infrastructure — tools for registering, discovering, and managing agents themselves (the operating overhead of automation)
Only ~3% genuinely new work — and most of it is about managing agents: selecting synthetic voices, developing AI personas, assessing agent trustworthiness

Why It Matters

ATE is a supply-side signal — what developers judge ready to automate, ahead of adoption data. The researchers connect it to labor economics: entry-level hiring in AI-exposed occupations has already fallen behind, and if routine tasks (how juniors traditionally learn) are automated, “how people become experts strikes us as one of the more important open questions about AI and work.”

The dataset is being released publicly so researchers, developers, and policymakers can track the agentic transition as it happens.

Frequently Asked Questions

What is the Agentic Task Ecosystem dataset?

The Agentic Task Ecosystem (ATE) is a Cohere Labs dataset aggregating roughly 696,000 published AI tools from 123,000 public MCP servers — the largest open dataset of its kind — used to measure which occupational tasks developers have built AI agents to fully automate.

What percentage of AI tools actually automate work tasks?

Only 2.6% of the 696,000 tools analyzed can carry out a recorded occupational task end-to-end, rather than merely informing a person or handling one step of a process.

Which occupations have no AI agent tools?

419 of 923 US occupations have no agentic tool activity at all, meaning developers have not built automation for nearly half of recognized occupations.

Does AI automation target routine or specialized work?

It depends on the field. In healthcare and computing, tools reach toward specialized tasks leaving humans the routine remainder. In legal, production, and sales, tools stay at the routine edges, leaving specialized work to humans. Specialized work resists automation when physical or interpersonal, and gives way when already conducted through software.