This week, at a developer event in London, OpenAI showed Dots, a proactive AI assistant that learns the user's goals, monitors incoming emails and work tools, and starts tasks on its own without waiting for a command. Unlike Codex, Dots does not wait for instructions. The product was first announced a month earlier at a developer conference in San Francisco, and since then the company has been presenting it at closed demonstrations, where the key question is how far the assistant can go in the user's work without explicit permission.
For now, Dots is unavailable in Europe: the company is negotiating with EU regulators. At first, OpenAI plans to offer the service to corporate clients and users of its premium Pro plan, and only in the United States. This choice of audience is linked to the fact that the product requires significant computing resources, and therefore significant costs, which the mass market is hardly able to cover yet without restrictions.
Competition in this market is fierce. The company's rivals include Muse from Meta, Grok Bot from xAI, and Duet from Anthropic. Muse reportedly enjoys popularity among mass consumers, precisely the segment where OpenAI is not yet betting. Elon Musk called the costs of running Grok Bot unacceptably high, which highlights a common problem in the industry: proactive agents are more expensive than ordinary chatbots because they constantly monitor data and carry out actions rather than answering single queries.
OpenAI also acknowledges that Dots require substantial computing resources, so some tasks are routed to the more compact Terra model instead of the large Astra model. According to available data, Terra is not inferior to Astra in some areas while costing ten times less. This approach makes it possible to reduce the cost of serving each request without giving up the powerful model entirely for complex tasks.
The financial side of the issue is also becoming noticeable to customers. A number of companies, including Apple and Goldman Sachs, reportedly intend to limit employees' spending on AI use. For a corporate buyer, this means that even a useful tool may run into an internal budget, and the manufacturer is forced to prove economic return, not just technical superiority.
Privacy presents a separate problem. Meredith Whittaker, president of the Signal messenger, warns of the risks of the assistant's deep access to data. In her words, contacts who never consented to having an AI read their messages are put at risk: their correspondence comes into the system's view without their knowledge. She also points to the danger for human rights activists in authoritarian states, for whom a leak or unauthorized access to correspondence could have serious consequences.
OpenAI proposes building trust gradually, starting with access to the calendar, and only then expanding the assistant's powers. Users can see which data Dot accesses, and the company is considering more detailed control settings. Whether these measures will prove sufficient for regulators and critics is not yet clear, and the question will depend on how the negotiation process in Europe goes.
The topic will also be discussed at the Fortune Brainstorm AI conference, which will take place December 7–8 in San Francisco.