OpenAI explains how dots became its always-on personal assistant

OpenAI’s Alexander Embiricos, who leads the company’s enterprise product team, told TechCrunch Disrupt attendees this month how its Codex coding tool evolved into dots, the always-on personal assistant OpenAI released at its late-September developer day. The assistant runs inside the ChatGPT desktop and mobile apps, available to subscribers on the $100-a-month Pro tier or higher; once connected to a user’s calendar, email and other tools, it works in the background to help get tasks done.

Embiricos said much of what dots does was already possible with Codex, OpenAI’s coding tool released earlier this year, but required far more hands-on setup. The shift, he said, was less about new technology than about framing the product so it feels intuitive — treating the assistant as a single ongoing entity rather than a series of separate chats. “There’s kind of this convergent evolution happening,” Embiricos said of the parallel push across the industry to build AI assistants.

That proactivity carries a cost: more token usage for OpenAI, and a more forward-leaning assistant for users. By Embiricos’ own account, the timing isn’t yet well-tuned — dots will flag calendar conflicts and deadlines, but not always with good lead time. He described an early version alerting him to a scheduling conflict three minutes before it happened, far too late to act, and said he expects the timing to improve as the product matures.

dots also runs a feature Embiricos called “proactive research”: on a set cadence, it reviews what it knows about a user and whatever accounts it can reach, then proposes ways to help. Meta’s Muse runs a similar process and has drawn criticism for the detailed profiles it builds of users’ relationships. OpenAI says dots inherits ChatGPT’s own privacy controls — users can disable model training and adjust which resources the assistant can reach — though those controls aren’t yet granular, and the app can push hard for access; it currently refuses read-only permissions for cloud drives. Embiricos said OpenAI eventually wants much finer-grained controls, but for now leans on the model’s own judgment, adding that the company expected pushback over its conservative, permission-heavy defaults and plans to adjust them over time.

Embiricos was skeptical about how far personal AI agents can scale: businesses and other people, he said, don’t necessarily want to deal with an AI agent, and he counts himself among them — he keeps a human assistant partly because OpenAI’s corporate partners expect to interact with a person, not a bot. OpenAI, he said, is moving cautiously on giving agents more real-world capability, including deeper work with payment processors, choosing to coordinate with the wider ecosystem rather than race ahead alone.

Why it matters

dots shows how major AI labs are converging on persistent, proactive assistants as the next consumer battleground — and how privacy controls and scaling, not raw model capability, are becoming the harder problem to solve.