Each ChatGPT Dot works from its own cloud computer after the user walks away

In Tom's Guide's account of the launch, a Dot might notice an invoice you still need to send, prepare it and put it in front of you for approval ^[9]. The person's part of that job comes down to one decision at the end.

By the same account, users of ChatGPT now type a prompt, wait for the answer, then carry it somewhere else by hand ^[1]. "Until now, even when ChatGPT gave me exactly what I wanted, I was usually the one copying the answer, opening another app and doing something with it," the Tom's Guide writer wrote ^[11]. Dots are pitched as giving ChatGPT "a place to actually do things for you" ^[12].

The worked examples stop short of the decision. The dinner Dot could build itemised cart estimates with local taxes and a tip allowance, then send the comparisons so the person picks one to order ^[8]. The invoice Dot stops at approval ^[9]. Two of the three examples in the coverage end with a person making the call ^[1].

The rollout decision has four parts, and the product answers each one differently. On where it runs, each Dot works inside an isolated cloud virtual machine and gets nothing on a user's laptop without explicit permission ^[3][4]. On what it can touch, the answer is whatever the user connects, from more than 4,000 plugins ^[5]. The isolation does not extend to connected accounts. A Dot carries out tasks using the services it is connected to ^[2], and those accounts are the part an admin has to answer for. On how it reports, OpenAI says Dots can be reached by voice call, Slack and Microsoft Teams, with SMS planned ^[6]. On how it starts, someone sets up each Dot on ChatGPT's desktop web or desktop app, choosing its purpose and connecting its services, before messaging it from a phone ^[7].

According to Tom's Guide, OpenAI does not expect users to spend all day watching their Dot work ^[14]. So a pilot gets judged on what comes back. The writer described the difference as "how much of the process happens after you've made the initial request" ^[13]. Count those steps per task. Then count how often the prepared work is approved without edits. The number of messages a Dot posts in Slack or Teams belongs in neither figure.

Sort errands on two axes. One axis asks whether the output stays inside the team or goes to someone outside it. The other asks whether a person can check it quickly at the approval step. Internal and quick to check, like the dinner comparison or a list of candidate clips from an interview transcript ^[8][10]: hand it off first. Internal and slow to check, like keeping tabs on a research topic ^[15]: hand it off and read a sample of the results. External and quick to check, like the invoice ^[9]: hand it off with approval required before anything is sent. External and slow to check: keep it with a person for now.

I'd start in the first box. The tradeoff is that the first box holds the least valuable errands. The invoice sits in the third, where the approval step is the only check before it leaves the building.