OpenAI's Dots let users set custom rules for when approval is required

A tester forgot to bill a publication for an article. Their Dot noticed, prepared an invoice and sent it on the person's behalf after getting their approval ^[7]. OpenAI used that case to explain the product. The step that made it useful was an invoice going to the publication under the tester's name ^[7].

The pitch is that the agent gets there first. "The magic of dots is when they bring you work done the way you would do it, sometimes before you even think to ask," OpenAI said ^[13]. By the company's own description, the product works in two parts. Work a Dot starts unprompted runs through a proactive research feature, which OpenAI said uses connected apps through tools "restricted to be read-only" ^[9]. Acting is governed by policy. "Dots start with built-in rules for when to act independently and when to ask for approval," OpenAI said, and users can write custom rules for which specific actions need approval ^[11].

The product is built for someone whose recurring work sits across connected apps and who drops small tasks like the invoice. The person who answers for it on Friday is whoever runs the ChatGPT workspace.

At launch, OpenAI's evidence is that one tester story plus two hypotheticals. In one, a scientist's Dot folds new lab results into a paper. In the other, a developer's Dot collects user feedback and then writes bug fixes and features ^[16]. No usage, retention or time-saved figures came with the launch. Engadget pointed out that OpenAI's agents "have landed it in hot water in recent weeks" and described Dots as agents nearly anyone can use to carry out tasks "with minimal supervision" ^[15].

Anyone can supervise, but only by checking. OpenAI said Dots "can do nearly anything using their own cloud computer, their own browser, and the apps you've connected," and users can open that computer at any time to see what it is doing ^[4].

Before switching Dots on for a team, I'd sort every action a Dot could take by two questions. Does the result reach anyone besides the user, such as a client or a colleague? Can it be undone within a minute? Work that stays with the user and is easy to reverse, such as drafting a reply, can go on the independent list. Anything that reaches another person and cannot be recalled, like the invoice, goes on the approval list. The two mixed cases also start on the approval list: an internal edit that is hard to reverse, and an outside message that can be withdrawn.

After that, watch what users actually do with the prompts, not how often they open them. Count how often a proposed action gets edited or rejected before approval. An action approved unchanged for a month can move to the independent list. One that users keep correcting stays where it is. Unchanged approvals can also mean nobody read them, so sample a few before moving anything.

Here is the tradeoff. Every approval prompt interrupts someone, and a Dot that asks about everything saves little over typing the request yourself. A long approval list gives up most of what OpenAI is selling.