OpenAI’s Dots is designed to keep working in the background, using connected tools such as a calendar, email and computer to offer help without waiting for a new prompt. The product puts a more approachable face on capabilities that could already be assembled with OpenAI’s coding tools, while raising practical questions about privacy, reliability and how far people want an assistant to act.
From coding tools to a persistent assistant
Dots is available in the ChatGPT desktop and mobile apps to subscribers on the Pro tier or higher, with plans starting at $100 a month. Once connected to a user’s resources, it aims to help with work by keeping a continuing picture of what is going on.
Alexander Embiricos, who leads OpenAI’s enterprise product team, described the design idea as making the assistant feel like one entity and one continuing conversation. That framing, he said, can make the experience more intuitive than managing many separate interactions.
Much of the underlying functionality, Embiricos said, could be achieved with the Codex app, OpenAI’s coding tool released at the beginning of this year. But doing so takes more setup and guidance. Dots packages that kind of capability for people who may not want to configure agents themselves or keep directing them.
The product name also reflects a goal of making the assistant feel like something that could eventually be ubiquitous. OpenAI has been moving from chat-based interaction toward tools that can take on tasks, and Dots extends that direction into an assistant intended to remain active over time.
Proactive help depends on getting the timing right
A persistent assistant is meant to notice opportunities to help without being prompted. OpenAI calls one part of this approach “proactive research”: Dots considers what it knows from conversations and the resources it can access, researches what might be useful, and proposes possible help.
That can make the assistant feel attentive, but the timing is not always useful. In testing described by TechCrunch, Dots could flag calendar conflicts and deadlines, though sometimes not early enough. Embiricos recalled an early alert that arrived three minutes before a conflict, leaving little time to respond.
He expects timing to improve. He also relayed an account from an engineer at a drone delivery startup: after an outage began, the engineer’s dot had noticed it, analyzed it and brought forward information that could help resolve it by the time the engineer became aware of the problem.
These examples point to the promise and the standard the product must meet. An assistant that notices a problem early can save a user effort; one that alerts them at the last moment may add pressure instead. For a background agent to be dependable, it needs to surface relevant information at a moment when the user can still act on it.
More access creates a privacy trade-off
Dots becomes more useful when it can draw on personal information, but connecting email, calendars and computer resources means users must decide what to make available. OpenAI says Dots uses the same privacy controls as ChatGPT itself: people can turn off model training and manage which resources the assistant can access.
Those controls are not yet granular in every case. The article reports that the app can be pushy about permissions and refuses read-only access to cloud drives. Embiricos said OpenAI needs to build much more detailed controls, while also pointing to the assistant’s intelligence and discretion as part of the balance.
There is a related question about how an assistant’s data use relates to its parent company. Embiricos described proactive research as a recurring process that considers what Dots knows about a user and what they have allowed it to access, then proposes ways to help. The article contrasts this with Hark, a startup that says it does not collect data to fuel an advertising business the way OpenAI and Meta do.
Permission choices are part of the product experience, too. Embiricos said early versions used conservative defaults and asked users for permission frequently, a response OpenAI expected to hear feedback about and planned to tune over time. More control could help users decide what the assistant can do and how much initiative to give it.
Agents also need the world to accommodate them
Even a capable personal assistant depends on other people and organizations being willing and able to deal with AI agents. Embiricos said business partners want to interact with a person, and he has a human assistant himself. He also said he does not intend to meet with someone else’s agent.
That friction matters if assistants are to handle tasks across businesses, websites and payment networks. OpenAI says it is working closely with the broader ecosystem as it adds ways for agents to interact with the world. Embiricos also said the company is in conversations with payment processors, without sharing further details.
Allowing an assistant to spend money would make the privacy and trust questions more consequential. The article’s author describes email access as already a significant ask and bank account access as off limits for now. Dots’ broader test is whether it can be proactive enough to help while giving people meaningful control over what it sees and does.