ChatGPT Work's real ask is your Slack, and somebody has to say yes on everyone's behalf

A harness, in the vocabulary of the engineers building these products, is the software wrapped around a model that decides what information it sees, what tools it can use, and how it presents answers back to you ^[6]. That is also a serviceable definition of an access policy. The open question for ChatGPT Work is who gets to write that policy, and on whose behalf.

The wager Andrew Ambrosino describes is coherent for someone whose job is testing this on himself: he accepts that a drafting agent may pull from a private DM, treats it as a personal hit he is willing to take, and reports it has not landed yet ^[5]. Copied into a team, the arithmetic changes owner. The messages an agent surfaces belong to colleagues, while the grant was made by whoever held the credentials ^[4]. Tibault Sottiaux, who runs core product work including Work, describes ChatGPT completing entire complicated tasks autonomously in a way that is "delightful and safe" ^[12]. In that sentence, "safe" is doing all of the load-bearing work, and none of the scoping.

The adoption spread in the OpenAI-backed study is usually read as a usability problem, and OpenAI reads it that way too: the tool was, in Ambrosino's words, actively hostile to the company's own comms and finance staff, and the team spent February onward making it general purpose ^[8]. The remaining distance between insiders and paying organisations is 81 percentage points, about 5.8 times the rate ^[13][14]. Now look at who the 98% are ^[7]. They are employed by the company receiving the data. Connecting a mailbox moves nothing across an organisational boundary for them. For an accountant or a doctor, the same click adds a counterparty.

Christian Catalini's warning on a16z's Time to Build blog is that if the labs cannot rapidly get hold of the key complementary assets needed to scale AI in the market, value accrues elsewhere ^[10]. In coding, that asset was the repository. Outside engineering, it is a pile of permission grants that belong to other people, and no amount of model quality wins one. Harvey in law and Clay in sales are chasing the same professional customers with a model-agnostic approach ^[11], which means they arrive at the identical consent screen with the identical problem.

What none of the reporting establishes is how ChatGPT Work behaves on partial access: whether a Slack connection scoped to three channels still completes multistep work, or only demonstrates it. That is the figure a team lead needs before signing, and it has not been published.