In an interview published October 26 on TechCrunch, Alexander Embiricos, head of product for background AI experiences at OpenAI, shared details about the development process of the personal assistant app dots. The project reflects an effort to bring agent capabilities that were once reserved only for programmers into the everyday work life of general users.
These remarks provide a realistic picture one month after the company introduced dots to the market in late September, clarifying the gap between automation expectations and real-world user experience.
From coding environments to proactive personal assistants
According to Embiricos, the idea of building dots originated when OpenAI engineers realized that the multi-step reasoning capabilities of coding tools could adapt well to non-technical office tasks. Instead of merely responding when asked, dots is designed to maintain continuous context throughout the workday, proactively checking schedules, reading notes, and suggesting actions without requiring user prompts.
However, practical deployment revealed subtle obstacles in terms of psychology and user interface. TechCrunch noted that dots could intelligently detect schedule conflicts or financial deadlines, but the timing of notifications often lacked intuition. The app warning users to prepare for a meeting just three minutes before it starts is often too late to take action, while reminding them too early becomes annoying.
In particular, Embiricos himself candidly shared that he feels uncomfortable having to meet online with someone else’s AI agent, emphasizing that human presence still carries irreplaceable social value, even as automated systems have demonstrated the ability to consume resources many times over what humans use in continuous tasks.
Permission barriers and payment integration
The next major challenge for background assistant models lies in security permission mechanisms. Testing by TechCrunch showed that dots currently requires full permission to edit and delete files on cloud storage services such as Google Drive or Microsoft OneDrive, refusing to operate if users only grant read-only access.
OpenAI acknowledged this as an experience bottleneck and stated it is researching more granular permission levels so users feel secure integrating personal data. In addition, the company is in discussions with intermediary payment gateways to allow dots to book tickets or pay for services on behalf of users within preset limits, similar to the direction rivals like Anthropic are pursuing with the Claude Duet line of proactive assistants.
Insights shared by the dots development team show that the greatest barrier facing AI agents today no longer lies in the model’s reasoning capabilities, but in learning to intervene at the right moment, respecting privacy, and building trust with users.