🔍 THE BOTTOM LINE: Don’t call the chatbot “Alex.” Don’t give it a Slack avatar, a job title, or a seat at the standup. New research from Boston University, surfaced by MIT Technology Review on June 29, 2026, shows that the moment a manager is told they’re “supervising an AI employee,” their error-catching collapses by 18% and their escalation rate jumps 44%. The “digital colleague” pitch being pushed by Nvidia, Microsoft, OpenAI, Anthropic, and Google isn’t a productivity story. It’s a liability offload dressed up as a culture fit.

Meet “Alex” — The Boston University Experiment That Should End the Hype

Boston University professor Emma Wiles ran the kind of study the agent-industrial complex would rather you didn’t see. Across 1,261 managers, participants were given the same AI output to review. The only variable was framing. Half were told they were overseeing an “AI employee.” The other half were told they were using an AI tool.

The “employee” group caught 18% fewer errors. They were also 44% more likely to escalate the work to a human manager rather than fix it themselves. Same AI. Same output. Same managers. The only thing that changed was whether the AI had a job description.

That gap is the entire story. Calling a language model an “employee” doesn’t make it more capable. It makes the human supervising it more passive, more deferential, and more willing to pass the buck up the org chart. Nobel laureate Daron Acemoglu has been warning for years that this is the trajectory — that “agents should be designed to augment, not replace” — and the BU data backs him up with hard numbers.

The Silicon Valley Sales Pitch Is Branding, Not Engineering

The reason this matters now is that the “digital colleague” framing has gone from keynote slide to product roadmap. Nvidia’s Jensen Huang talks about “digital humans.” Microsoft, OpenAI, Anthropic, and Google all market their agents as team members. We’ve covered the rollouts ourselves: Anthropic’s Anthropic puts Claude directly in your Slack channels as a proactive participant, and OpenAI’s OpenAI Launches Autonomous AI Workers for $20/Month — and They Work While You Sleep pitch positions agents as autonomous workers rather than tools you operate.

That’s not a coincidence. Anthropomorphism is the cheapest, fastest way to sell software. If a buyer believes they’re “hiring” an AI, they stop asking procurement questions and start asking HR questions. Annual contracts get longer. Switching costs get higher. And the moment something goes wrong, the vendor has a ready-made excuse: “well, your team was supervising it.”

Workers Don’t Actually Want the Job Cuts the Tech Bros Promise

Here’s the part the displacement panic usually buries. A separate Stanford-led study across 1,500 workers in 104 jobs found that employees do want automation — but in different places than the experts flagged as “most suitable for AI.” Workers actively resisted ceding the tasks involving judgement, context, and accountability. The Cloudflare Fires 1,100 Workers Because AI Agents Did Their Jobs — the Most Honest Layoff Yet narrative — that agents are coming for entire roles — runs directly against what the people actually doing those roles say they want.

This is the second-order finding the vendors don’t like. Workers can tell the difference between a tool that helps them ship faster and a tool that’s being installed to make them redundant. The first gets adopted. The second gets resisted, quietly, in every Kiwi and Australian office where someone stops volunteering ideas because they’ve clocked what’s actually going on.

The Accountability Gap Is a Feature, Not a Bug

Read the BU numbers again: 18% fewer errors caught, 44% more escalations. That’s not a productivity gain. That’s a workflow engineered to diffuse blame. If “Alex” makes a bad call, the manager who approved it can point at the AI. The vendor who sold it can point at the manager. The escalation lands on someone two rungs up the chain who has even less context. This is sophisticated blame dumping, sold to you as workforce transformation.

The legal exposure for NZ companies is the part nobody’s pricing in. Under the Health and Safety at Work Act, the person signing off the work is the person liable for it. You can’t delegate duty of care to a model that hallucinates. The “Alex did it” defence won’t survive a Worksafe investigation, and it certainly won’t survive an insurance assessor.

NZ Angle — Don’t Import the Theatre, Import the Tools

New Zealand workplaces are small enough that accountability still has a name attached to it, and that’s a competitive advantage we should be protecting, not trading away for a slide deck. Kiwi managers should treat agentic AI the way we treat any powerful tool with sharp edges: respect it, audit it, and never let anyone in the building forget that a human signed off on the output.

If you wouldn’t let a junior staffer send a client email unsupervised, you shouldn’t let an agent do it either — and you definitely shouldn’t give it a name, a Slack handle, and a place on the org chart. The framing is the attack surface. Strip it out and the rest is just engineering.

❓ FAQ

Q1: Is treating AI as a tool actually better than calling it an employee? A: According to the BU study, yes — and by a lot. The “tool” framing kept error detection 18% higher and cut unnecessary escalations by nearly half. The label changes the behaviour of the human, not the capability of the AI.

Q2: Are these agents really autonomous, or do they still need a human? A: They need a human. The 44% escalation rate in the BU study is essentially the industry admitting that someone has to catch what the model gets wrong. That’s not autonomy — that’s a junior with no training and no consequences.

Q3: Should NZ companies hold off on agent rollouts until regulation catches up? A: No, but don’t import the Silicon Valley framing either. Build internal guardrails around process failure, not personnel management. Audit the outputs, not the AI’s “performance.”

Q4: Where does Daron Acemoglu fit into this? A: His core argument — that AI should augment human skill rather than replace the worker — is exactly what the BU data validates. The “employee” framing is the opposite of augmentation. It’s substitution with a friendly face.

Q5: What’s the single safest habit a Kiwi manager can adopt this week? A: Stop calling it “Alex.” Call it “the model.” Force every review meeting to start with the human reviewer naming what they checked. The framing is the first line of defence.

🔍 THE BOTTOM LINE (Synthesis): The “digital colleague” pitch is corporate theatre designed to make you trust the software more, audit it less, and absorb the liability when it breaks. The Boston University research is unambiguous: the moment you call a model an employee, the humans around it get worse at their jobs and better at passing the problem upstairs. New Zealand’s edge in the agent era isn’t going to come from copying the Silicon Valley rebrand. It’s going to come from being the country that refused to pretend the chatbot had a job title.

📰 Sources

• AI agents are not your coworkers — MIT Technology Review (June 29, 2026)
• Daron Acemoglu — Wikipedia
• Boston University, Emma Wiles — manager framing study, 1,261 participants
• Stanford University — worker preferences study, 1,500 workers across 104 jobs