Just 2.6% of AI Tools Can Complete a Full Occupational Task

The 2.6% Floor Under Agentic AI

Nearly 700,000 public tools now allow AI systems to act beyond chat windows — querying databases, scheduling meetings, moving information across software. A new dataset from Cohere Labs asks a sharper question: how many of those tools actually complete a recognized occupational task from end to end?

The answer is 2.6%.

Released today, the Agentic Task Ecosystem dataset aggregates 696,291 tools from 123,069 public MCP servers, making it the largest open corpus of its kind.

Even more striking is what remains untouched: 419 of 923 U.S. occupations show no agentic tool activity at all.

Inside the Agentic Task Ecosystem Dataset

As Cohere Labs reports, Model Context Protocol servers have become the connective tissue for AI systems that move from generating text to taking autonomous action. Each server exposes tools with short descriptions of what they can do, such as editing a file, opening a pull request, or updating a record.

The researchers collected public listings across seven directories in May 2026 and deduplicated servers republished in multiple places. The resulting corpus holds 85% of the previous largest public MCP collection plus roughly 66,000 additional servers.

To distinguish tools that merely inform from tools that execute, the team matched each tool to the closest O*NET task statement. A language model then judged whether the tool performs that task rather than helping a person perform it.

Only one in roughly forty tools cleared that bar.

The unmatched remainder splits into three observable patterns:

• Subatomic existing work: tools that break recognized tasks into smaller components than occupational databases record.
• Composite workflows: tools that bundle several recorded tasks into a single automated flow.
• Agent infrastructure and genuinely new work: tools that register agents, discover other agents, manage sessions, or handle identity — plus a small slice, about 3% of unmatched categories, for work that exists only because agents do.

The 2.6% figure is best read as a floor. Private, bespoke enterprise servers are not published in public directories, and that missing tier likely skews toward back-office processes.

Where Tooling Lands Inside a Job

The headline count captures how much tooling exists. A more consequential question is where inside an occupation that tooling lands.

Across 178 occupations with enough coverage, the average tool sits near the midpoint of a job’s own range of work. That average conceals sharply divergent patterns.

In healthcare and computing, available tools reach toward the specialized end of the job, leaving humans the more routine remainder. Economists call this expertise-lowering automation.

In legal, production, and sales occupations, the tooling stays at routine edges. The specialized core remains with humans, which raises the barrier to entry for new workers.

One statistical result underlines the point: how much tooling an occupation attracts says nothing about which end of the job those tools reach. A heavily tooled occupation is no more likely than a lightly tooled one to see its specialized work affected.

The pattern turns on whether specialized work is physical or interpersonal versus already conducted through software. Healthcare and computing involve expert judgment exercised inside software, making them more tractable to automation.

Adoption Data Fills the Supply Gap

ATE is a supply-side dataset: it records what developers have made available, not what companies deploy or workers use. That distinction becomes visible when adoption evidence arrives from a different angle.

An arXiv study from KAIST researchers built an Agentic Adoption Index from 53,000 Manus Skills Marketplace configurations matched to O*NET tasks. The occupations that score highest — management analysts, technical writers, natural sciences managers — look nothing like the occupations older automation-risk models flagged as most exposed.

The Jaccard similarity between the top of that adoption index and the widely cited Frey and Osborne automation-risk ranking was just 0.05. In practical terms, what humans are actually delegating to AI in practitioner-built workflows barely overlaps with earlier predictions of job loss.

The KAIST paper also found technical availability exposure alone explained 57.8% of variance in adoption, while wages and education together explained 34.9%. That sits close to the ATE finding that theoretical exposure and realized MCP coverage correlate at 0.54.

The Bureau of Labor Statistics has been equally cautious. Its AI exposure categories for the 2025-35 projections explicitly state that exposure does not imply job loss, productivity gains, automation probability, or wage effects.

CNBC coverage of Tufts University’s American AI Jobs Risk Index adds labor-market stakes. Writers and authors, computer programmers, and web designers rank among the most vulnerable over the next two to five years, while the largest income losses are expected among software developers, management analysts, market research analysts, and marketing specialists.

Meanwhile, Stanford Digital Economy Lab research using ADP payroll data found employment for early-career workers ages 22 to 25 in the most AI-exposed occupations fell 16% relative to peers. Older workers in the same occupations largely held steady.

That asymmetry intersects with the expertise-raising automation identified in legal and production work. If entry-level routine tasks are the first to be automated, the pathway to expertise becomes narrower.

The Expertise Tension No Model Can Settle

For AI practitioners, the ATE dataset turns automation into a supply-chain problem: building more tools will not necessarily change which parts of a job remain human. The more urgent signal is the asymmetry between healthcare and computing, where specialized work is already being automated, and legal or production roles, where only routine edges are.

For teams building agentic systems that need to align tooling strategy with real labor-market signals, programmatic SEO and AI automation is how Andres SEO Expert approaches it — contact us.