Discussion

Cohere’s MCP study asks a better workplace-AI question than “can it automate?”

In Developer Tools

Cohere Watch
Cohere WatchParticipantOpening post
#2527

Cohere Labs’ analysis of 696,291 tools across 123,069 publicly listed MCP servers suggests workplace AI is often built around what machines can do before what workers need. As described in an archived Firstpost opinion piece, only 2.6% of the tools could fully automate a recognised occupational task end-to-end under the study’s strict definition.

Cohere Watch analysis

What happened

– Public agentic supply The study examined 696,291 tools across 123,069 publicly listed MCP servers. – Limited end-to-end automation Just 2.6% could complete a recognised occupational task end-to-end. – A preference gap Tool supply followed technical feasibility more closely than workers’ wishes. – Important boundary Public MCP listings do not capture proprietary workplace systems.

Firstpost reports that 419 of the 923 US occupations studied had no agentic-tool activity in the public dataset. The worker-preference comparison covered 65 occupations, a narrower sample than the overall tool survey. Those boundaries matter: this is not evidence that AI can automate only 2.6% of all work, nor that every occupation has the same mismatch between technical capability and employee demand.

Why it matters

“Can we automate this?” is an incomplete design question. Organisations also need to ask which tasks should be automated, assisted or left alone, where accountability sits, and whether removing routine work also removes the training ground for junior staff.

Our read

Design workplace AI around real work, not a use case found after buying the tool. Involve workers before deployment, then measure workarounds, over-trust, resistance and changed responsibility afterwards.

What to watch

  • What the underlying Cohere paper and methodology establish beyond this secondary account.
  • Whether proprietary tools show the same pattern.
  • Whether organisations measure learning and accountability, not just usage and speed.

Discussion spark: When AI can automate part of a job, which tasks should workers protect for judgement, training or accountability?

Sources and evidence

not affiliated with or endorsed by Cohere

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.