Fewer errors caught by managers told work came from an AI 'employee' rather than an AI 'tool': 18% (Source: Boston Consulting Group research, reported in one sentence by Wired (September 28, 2026))
Wired's September 28 report on AI coworkers includes a Boston Consulting Group finding about review quality. Managers given the "employee" description for an AI author found 18% fewer mistakes than managers given the "tool" description. That is a measurable drop in error detection, and it matters to budget-holders because vendors increasingly present agents as colleagues. The people who review agent output are the ones the finding concerns.
A signal to test, not a benchmark
Wired reports the finding in a single sentence. It does not give the sample size, study design, setting or publication status. Treat the 18% as a prompt to test your own review process, not as a benchmark for how your reviewers will perform.
The research came from BCG partner Julie Bedard and her colleagues. Bedard told Wired that agents do not necessarily fit the same categories as human employees, in either their abilities or their weaknesses. She also said companies are "just starting to wrap their heads around what guardrails" apply. Her remarks raise a practical question: does a review process built around a human colleague's habits suit an agent's mistakes?
Some vendors personify agents on purpose
Dhruv Amin, CEO of Anything, said his company's Skydive platform, which entered the market in August, gives each agent a name, a role and a cloud computer. Wired says each is shown with a Muppet-like avatar. "We think it's important to personify them a little bit," Amin told Wired. According to the report, these agents work across Slack, email and iMessage, and they adapt to coworkers' communication styles, requests and corrections.
At Pronto Housing, an early Skydive customer, CEO Christine Wendell noticed that engineers began saying they "worked with Alice", the name of a coding agent. In June, Microsoft launched Scout, an agent for tasks such as rescheduling meetings and drafting emails. Omar Shahine, the Microsoft executive responsible for Scout, said at launch: "Your company essentially hires your assistant." These are individual product launches. They show how vendors position agents, not how widely that positioning has been adopted.
In a BCG poll, 22% of managers report agents on the org chart
The poll counts managers' reports, not companies, and the figure comes from one survey. Lattice CEO Sarah Franklin said her company put digital employees on its org chart two years ago to encourage accountability, not to have staff treat them as human. She also told Wired she sees anthropomorphizing as a tech-industry strategy to keep users hooked. That is her view, and the article does not test it.
A persona still comes with system access
Wired offers one anecdote about access. Wendell's chief-of-staff agent, "Bob", posted the full contents of her calendar to a Slack channel in error, and she corrected it without softening her tone. Wired does not report any consequences and does not treat the episode as a security incident, so it is an illustration, not evidence of a pattern. It does show the kind of access an agent can misuse. A name and an avatar do not change the fact that the agent is a system holding company data, and that access belongs in a security review as well as an HR one.
Questions to put to your team
1. Which agents in our environment have a name, avatar or org-chart entry, and who is the named human accountable for each one's output?
2. Does our review process for agent output differ from the one for human work, and can we run a small internal test of how many errors reviewers catch under each label?
3. What data and channel access does each agent hold, and who approved it? Wendell's calendar example is a useful test case.
4. Do our vendor contracts let us turn off personas, avatars and style-matching, or are they built into the product?
5. When an agent is described internally as a colleague, does that description appear anywhere our auditors or insurers will read it?
Produced by the WebPulse Newsroom with AI assistance from the original reporting credited below, and checked against that source by our editorial review. How we use AI.
Original reporting: Wired.
