An assistant that keeps working after you close the window
OpenAI introduced Dots on 29 September 2026: assistants in ChatGPT that run continuously and carry on with tasks even when nobody is asking a question. According to Vellum, a Dot is a named, persistent assistant based on GPT-6 Astra. TechCrunch quotes OpenAI's description: "always-on agents". An agent here is an AI system that carries out work steps itself instead of only producing text.
This first part of a series looks at what is new about Dots, where the limits are and what the situation is in the EU. All information comes from media reports and the vendor's own account. We have not tested Dots ourselves. Figures that come from OpenAI itself are marked as such.
What is actually new
The novelty lies in how it works rather than in the model. A Dot runs continuously and has its own workstation.
Each Dot works on its own computer with a browser in the vendor's cloud, not on your device. According to Engadget, users can open this computer at any time and see what the Dot is doing. Access to your own laptop is optional and starts switched off (Vellum).
A Dot can also do research on its own initiative. This research is explicitly limited to reading: it sends nothing, changes no content in connected apps and controls neither a browser nor a computer (Vellum). Then there are scheduled tasks that run to a timetable without anyone triggering them. And it can be reached where people work: via Slack, Teams and other platforms, with SMS to follow (TechCrunch). Plugins can connect it to more than 4,000 apps (Engadget).
A chat assistant waits for your next question. A Dot keeps working, and that changes the job for decision-makers. How good the answers are becomes secondary. What matters more is how long an assistant runs, which systems it can access and how far it may act on its own.
How far a Dot may act on its own
OpenAI handles this on several levels. Built-in rules determine when a Dot acts independently. With their own rules ("custom rules"), users specify which actions need their approval first (Engadget).
According to Vellum's overview, tasks you trigger and scheduled tasks can run under separate rules and approvals. A weekly report that a Dot produces by itself every Monday therefore doesn't need the same latitude as a follow-up question in an ongoing conversation. Vellum lists further safeguards. Actions that affect accounts or share information are checked automatically. Monitoring can pause suspicious work. Tasks such as changing a password stay with the user.
These are rules the vendor designed. How well they hold up in daily use is an open question: Vellum itself notes that real-world error rates have not yet been determined independently.
Oversight is work: OpenAI's own figure
The most revealing figure on Dots comes from OpenAI itself. In a controlled test, the company checked how often a Dot steps outside its remit. We know the result second-hand: The Rundown reports on the test.
The variable was how many other tasks came up between the instruction and its execution. With five intervening tasks, the share of flagged moderate scope violations was 8.6%. With ten, it rose to 19.7%. According to this report, no serious violations or data leaks were observed in the same test.
A second indication comes from NBC News: OpenAI held back the GPT-6.1 Astra model because it did not quite clear the bar for staying within task scope and permissions.
Our reading, which does not follow from the test itself: the more that happens between instruction and execution, the more closely someone has to look. Oversight costs working time. If you plan for an assistant that runs continuously, you also plan for the person who checks its work. So approval rules are not something you set once and forget. Three questions help:
- Who approves? One named person per area of responsibility, with time set aside for it.
- How much runs in between? This is also a variable you control yourself.
- When do you review the rules? At the latest when an assistant gets additional tasks or access.
The situation in the EU
According to the available reports, Dots is not available in Europe at launch. Vellum says OpenAI is rolling out market by market; for ChatGPT Pro, the European Economic Area (EEA), the United Kingdom and Switzerland are excluded at launch. NBC News reports that Dots is not available in Europe or the UK.
For Business Premium the situation is unclear. This plan also includes the first Dot (Vellum). The sources do not mention an explicit regional exclusion for it. TechCrunch writes of "eligible markets" without listing them. Whether a Business Premium subscription in Germany includes Dots is therefore open.
For Enterprise, Education and Healthcare workspaces there is a beta that an administrator switches on (Vellum). It is off by default (The Rundown).
Once Dots is available in your market, your vendor should answer three questions before you approve anything:
- Region: In which countries is the plan you use enabled, and does that also apply to the beta in your workspace?
- Data storage: Where are conversations, the Dot's memory and the content from connected apps stored? According to Vellum, what the Dot has already learned is kept when you disconnect an app; only a reset deletes conversations, memory and scheduled tasks together.
- Logs: What records exist of what a Dot has done, and who can view them?
The answers do not replace a legal review. They do provide the facts such a review needs.
What to do now and what comes next in this series
For most companies in the DACH region, the next step for now is to watch. You can put the waiting time to use. Take a few recurring processes, such as preparing quotes or maintaining customer data in the CRM. For each step, decide what an assistant may do: only read, make a suggestion, or act after approval. Note who approves and whether the task would run on request or on a schedule. If you write these limits down today, they will be ready when an agent becomes available, whichever vendor it comes from. The 3-tier model for AI agents classifies tasks by risk. The AI Diagnostic clarifies where your company stands overall.
The series continues:
- Part 2 compares hosted agents such as Dots with a self-built agent harness: a company's own environment in which it runs a language model with its own tools, rules and logs.
- Part 3 asks what this means for AI transformation in mid-sized companies: responsibilities, approval levels and the path from experimenting to operations.
- Part 4 turns the perspective around: what changes when your customers send their own agent to make enquiries, compare offers or place orders with you?
Which task in your company would you hand to a continuously running assistant today, and who would approve it?
Frequently asked questions
Is Dots available in Germany? Not for ChatGPT Pro: according to Vellum, Pro availability at launch excludes the EEA, the UK and Switzerland. For Business Premium the situation remains unclear. Ask the vendor directly about the status for your contract.
How is a Dot different from a normal ChatGPT chat? A chat answers while you are using it. A Dot is a named assistant with its own cloud computer that also handles scheduled tasks and can be reached via Slack or Teams.
Can a Dot send messages without asking first? Not during proactive research, which is strictly read-only. For tasks you trigger or schedule, built-in and custom rules decide when approval is needed. If outgoing messages are a sensitive point for you, make sending an action that requires approval before a Dot starts working day to day.