BCG's large scale experiment finds that framing AI as a colleague or employee erodes the human oversight that's supposed to keep AI outputs reliable — and points to a different organizational frame that works.
BCG's large-scale experiment finds that framing AI as a colleague or employee erodes the human oversight that's supposed to keep AI outputs reliable — and points to a different organizational frame that works.
When BCG put AI agents on a simulated org chart, error detection dropped. Not because the AI got worse — because the human did. The "employee" framing doesn't just rename the tool; it triggers a specific behavioral chain that erodes the oversight that's supposed to keep AI outputs reliable.
A study published in Harvard Business Review on May 6, 2026, led by BCG's Matthew Kropp with co-authors Bedard, Wiles, Hsu, and Krayer, tested exactly this. Across more than 1,200 HR and finance professionals in the U.S., Canada, and the EU, the team ran a controlled experiment in which the same error-laden workplace document was attributed to three different sources: a human employee, an AI tool, and a named AI "employee." Participants reviewing documents attributed to the named AI "employee" identified meaningfully fewer errors than those in the other two conditions — and, according to the HBR summary of the BCG working paper, the gap is reported as an 18% reduction in error detection.
That number deserves a caveat. The HBR article body is gated for non-subscribers, and the BCG PDF linked from the study was not text-extractable in the version available for this reporting — so the 18% figure is being cited as HBR's reporting of the underlying experiment, not independently re-verified here. The qualitative finding, however, is consistent across both the HBR piece and Fortune's May 28, 2026 write-up: framing AI as an employee makes humans worse at catching its mistakes.
The effect is not really about AI. It's about what naming does to a human reviewer.
In the "AI employee" condition, participants reported lower felt accountability, were more likely to shift blame onto the AI when something went wrong, and were more likely to route the document to a human colleague downstream for another look. That last behavior is the most important: it doesn't eliminate human review, it just moves it one step away from the work itself, where the reviewer has less context and the buck has another place to land.
Kropp's explanation, as reported by HBR, is that anthropomorphization creates a "pass the buck" dynamic. AI cannot actually be accountable — it cannot be fired, demoted, or held to a performance review in any meaningful sense. So when a human stops scrutinizing AI output, there is no one left holding responsibility. A named human owner has to be assigned explicitly, or accountability is silently diffused.
This is what the data is measuring: the silent part.
The study lands in a workplace that has already moved past the question. According to the same HBR piece, nearly one-third of managers now frame AI as a teammate or employee, and more than 20% list AI agents on their company's work charts. Lattice, the HR-software vendor, drew the same line in 2024 when it announced "AI employees" as a category — and then partially walked the framing back after pushback from customers who weren't ready to put software on the org chart.
Cognizant's own research, cited in the same coverage, finds that 93% of jobs are already being disrupted by AI — six years ahead of the firm's 2023 projection — and that the productivity "activation gap" is the real bottleneck. In other words: the question is no longer whether AI gets used, it's whether organizations are set up to use it well.
The disagreement is now on the record at the highest levels. At Fortune's COO Summit on June 2, 2026, the panel could not reach consensus. Okta COO Kelleher — who has publicly named his agents Leo, Sloan, Hank, and Walker and argued that companies are "in denial" about redesigning work around AI — made the case that the colleague framing is the catalyst that unlocks adoption. Cisco's Katsoudas publicly rejected the frame.
The BCG data gives that disagreement a referee. The colleague framing is not a free choice with no cost. It measurably shifts human behavior in a direction that reduces the very oversight that makes AI output trustworthy in production.
The HBR piece doesn't just diagnose a problem; it points to a different frame. Instead of "AI employee," the authors and Kropp recommend treating AI as one of three alternative roles: a tool (with a human operator who owns the output), a teammate (a named role on a project without org-chart placement or performance review), or a junior reviewer (an AI whose work is checked by a more senior human, the way a junior's work is checked before it ships).
What changes in practice:
Each of these is a different answer to the same question: who owns the line when the AI is wrong?
The single change a manager, founder, or team lead can make on Monday: name a human owner for every AI-produced output that ships to a customer, a regulator, an internal stakeholder, or a downstream system. The owner is the person whose performance review absorbs the error, full stop. Not "we used AI for this." A name.
That is what the BCG data, the HBR reporting, and the live executive disagreement all point to: the choice is not between using AI and not using AI. It is between an org chart that silently diffuses accountability and one that doesn't. The 18% gap, the "pass the buck" mechanism, the Kelleher–Katsoudas split — they are all the same finding in different costumes.
The frame is the work.