On September 29, OpenAI introduced Dots, a new kind of persistent assistant that can work across several projects, use connected apps, and come back when something needs review. The rollout has started for Pro and Business Premium in eligible markets, while organizations can try an admin-enabled beta.
The interesting part is not simply that AI can do more in the background. It is that we need to become better at delegating. A vague instruction such as “keep an eye on the project” leaves almost everything to the model: what matters, when to act, and what it may change. A clearer assignment is both more useful and easier to supervise.
1. Outcome: Describe what should exist when the work is done. “Every Friday, give me a short list of decisions, blockers, and next steps.”
2. Sources: State where it may look and which material should carry the most weight.
3. Boundaries: Say what it may do independently and what it should only propose.
4. Review point: Decide when you want to inspect the work and which changes must always wait for approval.
OpenAI also describes safeguards including separate permissions, an activity view, and rules for when the assistant must ask before acting. That is a sound principle regardless of tool: start with read-only access, a narrow responsibility, and a clearly defined deliverable. Expand only after the results have become predictable.
This makes AI feel less magical, but much more useful. Think of an energetic trainee with endless patience: provide context, demonstrate what “good” looks like, and keep decisions with real consequences in human hands.