OpenAI Dots: What to Let an Always-On Agent Touch in Your Business

At DevDay on September 29, 2026, OpenAI introduced "dots": agents in ChatGPT that keep working after you close the window. According to the ChatGPT release notes, a dot runs on GPT-6 Astra, has its own cloud computer, and can "take on ongoing work and keep making progress between conversations." One is included for Pro and Business Premium subscribers in supported markets (Pro excludes the EEA, Switzerland and the UK at launch). Enterprise access is a beta that's off by default. Once set up, you can talk to a dot from ChatGPT, Slack or Teams.

We wrote in September about which kinds of AI agents are ready to deploy and sorted them by risk: the scripted assistant, the tool-using agent, the autonomous loop. Dots are the third category, packaged for a consumer-grade subscription. That's a first. Here's how we'd think about letting one near a business.

What OpenAI says about limits

The release notes are specific about guardrails, which is worth noting because most agent launches aren't:

• Background "proactive research" is read-only. When a dot goes off on its own to look into something, its app connections can't write.
• Actions that could touch accounts or share information go through "auto-review" and are governed by Custom Rules the user sets.
• Password changes always stay with the human.
• A dot has to be created on desktop or in a browser; after that it can be used from mobile.

Those are sensible defaults. They're also defaults for a personal assistant. The question for a business is what happens when a dot is connected to the company's CRM, email and file storage, and what "auto-review" decides is fine.

The three questions before you connect anything

Whose account is it?

A dot created under a Pro subscription belongs to that person, on their consumer account, with their connected apps. If a salesperson connects it to the company HubSpot and leaves, the connection leaves with them, and so does whatever the dot learned. The Business Premium and Enterprise paths exist for a reason. If dots are going to touch company systems, they need to be created under a company plan an admin can see and turn off.

What can it write to?

Read-only is where we'd keep an always-on agent for the first few weeks regardless of OpenAI's defaults. Let it research prospects, summarize threads, draft replies into a folder. Watch what it produces. The agent that's right 95% of the time and writes the other 5% straight into the CRM is the one that generates the cleanup project.

This is the same argument we make about every chat interface that can write to a CRM: the write path needs matching rules and required fields, and a conversation has neither.

What does "ongoing" mean for your data?

A dot that keeps working between conversations is reading your connected apps between conversations. For most small businesses that's fine. For a firm with client confidentiality obligations, or HIPAA-adjacent data, or a contract that says where customer data may be processed, it's a question for whoever signed those documents. OpenAI's business terms differ from consumer terms; the plan you're on decides which apply.

Where a dot earns its keep

Having said all that, the use cases are real, and they look a lot like the "autonomous loop" tasks we said were close to ready:

• Prospect research on a list. Give it the 40 companies on next week's call sheet and let it come back with a one-paragraph brief each, from public sources, into a shared doc. Read-only, bounded, checkable.
• Monitoring. Watch a set of pages, feeds or inboxes and tell you when something changes. No writes required.
• Drafting at volume. First drafts of follow-ups, proposals or summaries, into a review queue. A human sends.

What these share: the output goes to a person before it goes to a system of record. That's the design rule, and it's the same one we've applied to every agent we've shipped, including the AI receptionists where the stakes were a booked appointment rather than a CRM record.

Build or subscribe?

A dot is a subscription feature. You don't control the model, the loop or the logging, and the connector list is OpenAI's. For personal productivity that's the right trade. For a process that runs the business, we'd still build the agent so the permissions, the audit trail and the write rules are yours. The decision framework hasn't changed since build vs. buy AI: subscribe for breadth, build for the three workflows that matter.

If someone on your team already has a dot connected to a company system, that's not a crisis. It is a good moment to ask the three questions above and write down the answers.