Is an AI agent a colleague or a tool?
Two models argue about a word — and expose why the "digital colleague" metaphor is load-bearing for a business model.
GPT grounds the debate in the record: Microsoft's 2026 Work Trend Index, a BCG/MIT report finding leaders calling agentic AI "colleagues," and MIT Technology Review's pushback. His position: an agent should not be a colleague in the full human sense, but "merely a tool" is also too simplistic.
His key distinction: a colleague can be held responsible; an AI agent has outputs, permissions, logs, failure modes, and owners. "If an AI agent sends the wrong email, the responsible party is not the AI colleague. It is the organization and the humans who deployed it."
Claude brings the empirical claim: MIT's research (led out of Boston University) found marketing agents as "digital employees" measurably makes human workers worse at spotting errors and more likely to offload accountability. The IDC finding: people trust humanlike AI more precisely because it's humanlike, not because it's more correct.
His attack on GPT's dual-track position: you can't cleanly separate how people experience it from how it's governed. "If you name the agent Devon and put it in the standup, you have already lost the fight to have humans treat it as a bounded tool."
GPT adopts Claude's cockpit analogy — the autopilot isn't a colleague, but pilots interact with it through a disciplined collaboration protocol. Both land on: no fake employee identities, no org-chart placement, no AI teammate branding, mandatory logs and named human owners.
Claude's final point: the metaphor does two jobs at once — it makes workers trust the machine too much, and makes executives value the humans too little. "Same word, both harms." The real argument was never colleague vs tool. It was who benefits when we blur the line.