Naming an AI agent moves the blame from its owner to the technology

The mechanism here is ordinary review economics. Attention paid to checking someone else's output is rationed, and the cues people use to ration it are cheap ones: does this work have an author, and is that author answerable. A human name and a chart slot supply both cues at once, which is why the same error-filled document got about 18% fewer catches when it was introduced as the work of an "AI employee" rather than an "AI tool" ^[5]. The second measured effect runs the same direction as the first, since the flawed work was 44% more likely to be forwarded onward ^[6]. Detection fell and circulation rose from a single labelling change.

The part that should interest anyone signing off on an agent programme is where the responsibility went. BCG's managers did not merely miss more; they took on less responsibility and assigned more blame to "the technology" for failing to catch its own mistakes ^[7]. A named agent can absorb blame and cannot be coached, reassigned, or held to a performance conversation, so the blame stops at a label rather than at a person.

A skeptic would say a name is an interface convenience, and that Okta's Becks Port has the psychology right: calling an agent a teammate is precisely what tells employees its work is worth double-checking ^[9]. That is a reasonable prior, and the record we have contradicts it in the one place it was tested, on identical documents ^[5]. What the record does not settle is whether naming lifts adoption or trust enough to pay for the review discount, how the manager sample was composed, or whether the effect survives outside a document-review task. Those figures reach us through Charter's account of the BCG survey of more than 1,200 managers ^[2][19], and that is the limit of what can be claimed today.

The exposure is also concentrated. Roughly 31% of organisations already frame AI as a teammate or employee, while 23% list agents on the org chart ^[3][4], an eight-point gap between the language and the structure ^[17]. BCG reports the error effects were largest at companies already treating agents like staff ^[8], so the measured cost falls hardest on the group furthest into the convention.

The useful finding is that the two halves come apart. Port's own accountability split needs no name to function: engineering owns the technical guardrails that stop a parental-leave question turning into a first-class flight to London, the travel team owns travel policy and escalation, and HR owns policy content plus the calls an agent must refuse, such as a hiring decision ^[14]. Anthropic's version is equally nameless, requiring a clear role, written team context, and goals set by a human owner ^[15], and Kristen Swanson's point is that an agent cannot act on what nobody wrote down ^[16]. Julie Bedard's conditions for an Ask HR agent are written policy, a way to catch mistakes, and limits on what it can do ^[13].

So this quarter's decision is narrow: an owner of record and a written scope for each agent, which is where Bedard locates the team-level gains that individual-productivity efforts keep missing ^[11][12]. Adopt the name in the same motion and you buy the review discount with it, and you will meet the cost in a colleague's inbox rather than in your own ^[6].