ChatGPT Now Has Dots, and What's Interesting Isn't How Smart It Is but Who Holds the Access

On Tuesday, September 29, 2026, at DevDay, OpenAI announced Dots. In short, it's an always-on ChatGPT agent: it has its own cloud computer, its own memory, and keeps working even after you close ChatGPT. There's plenty of coverage of the features already. Here I'm taking the angle I think is more useful for a business owner: not how smart Dots is, but how OpenAI handles its access, approvals, and memory, because that's where your company will have to deal with it.

Dots in one paragraph

According to OpenAI's announcement, a dot is powered by GPT-6 Astra, works on its own cloud computer and browser, learns from feedback, and can connect to more than 4,000 apps through plugins. You reach it through ChatGPT, Slack, Teams, or a voice call (text messaging is coming), and you can open its computer at any time to inspect its work. The first dot is included in the Pro and Business Premium plans at no extra cost. Enterprise can try the beta only when the workspace admin turns it on. For Pro, OpenAI says the rollout is outside the European Economic Area, Switzerland, and the UK, while Business Premium is available in all supported ChatGPT regions. The rollout is gradual, so your account may not have it yet.

On the "own computer" capability in everyday work, I covered its predecessor in GPT-6 Astra for Non-Technical Professionals: Actually Useful, or Just Hype?. And on when a business really needs an agent that acts on its own, see Agentic AI: When You Actually Need It, and Platforms You Can Run Directly on a VPS.

First angle, OpenAI splits a dot's actions into three levels

The part I find most interesting is in OpenAI's help pages. In the Dots privacy, security, and safety FAQ, a dot's actions are split in a way that resembles the three levels of human oversight I covered in At What Point Should Humans Still Check AI and Automation Results:

OpenAI's own example sits right at level two: an early tester forgot to invoice a publication, his dot prepared the invoice, and sent it after he approved.

On top of those three levels there are other layers. Custom Rules let you tell a dot never to send emails, for example. Auto-review checks certain actions before they run, such as checking the recipient and content of an email against your instructions. Activity View shows what the dot is doing and has done. Custom Rules can't switch off the core safety requirements. I think this design follows the same principle: human oversight is set by the risk of the action, not by feeling.

But OpenAI itself writes that a dot can still make mistakes and that consequential work should be checked. Its protections against malicious instructions sneaking in through the web or email (prompt injection) are also described as reducing the risk, not eliminating it.

Second angle, a dot's memory sticks to one person

A dot learns from conversations, from its own notes about your preferences and decisions, and from connected apps. OpenAI's help pages have a few details I think business owners should read:

• A dot's memory can be shared with ChatGPT memory, and conversations with the dot can feed ChatGPT memory. Turning Memory off stops the sharing, but doesn't delete what was already received.
• You can't view or edit what the dot remembers one item at a time. The way to delete its memory is to delete the dot.
• Disconnecting an app doesn't erase what the dot already learned from it.
• Only the dot's owner can direct it. DMs or mentions from other people don't start work.
• On Enterprise, dot access is off by default. To remove someone's access, you have to review every role that grants it, and app authorizations and website sessions are managed separately (admin guide).

If you've read When the Company's AI Resigns Along With the Employee, you can see why this matters. If a dot works with an employee for months, it accumulates a lot of working context: decisions, preferences, the order things get done. That context lives in that person's dot, and on the pages I read there's no feature for an admin to browse or sort through a member's dot memory. It's a sharper version of the personal AI account problem I cover in A Simple SOP to Keep Company AI a Team Asset, Not Personal Property.

Third angle, an agent with its own identity

The most interesting part for business isn't the personal dot but what OpenAI calls specialist dots. According to the announcement, a company sets up each specialist dot with its own identity, credentials, and access to the systems it needs, so it can take on clear responsibilities inside the organization. OpenAI mentions lessons from internal testing in procurement, invoice processing, email marketing, customer support, and commercial contracting. It starts with enterprise pilots, where OpenAI's engineers work directly with an organization to define the dot's responsibilities, the tools it can use, and how people review and approve its work. OpenAI also says it's working with Microsoft so specialist dots can be managed through Agent 365.

When you think about it, that means AI agents are starting to be treated like team members: their own account, their own access, and a job description. And the work OpenAI mentions first isn't choosing a model but defining responsibilities, access, and how results get reviewed. That's organizational work, not model work. So the question for your company shifts too, from "which AI is smartest?" to these six questions (my suggestions, applicable to any agent):

1. What is its job, and what isn't? Write it like a job description.
2. What access does it have, and is that the minimum? Keep it separate from any one person's access.
3. Who approves risky actions? Decide which are level 2 and which are level 3.
4. Where is its memory stored, and who owns it? Keep important knowledge in documents the company owns.
5. Who is accountable when it's wrong? The company is still responsible for what its agent says and does.
6. What does offboarding look like? Revoke access, accounts, and sessions, and make sure the knowledge doesn't leave with it.

For comparison, the internal assistant I built in Check the Team's Work and Daily Tasks With Just a WhatsApp Chat is far narrower: it only answers questions from the data in Google Sheets, over WhatsApp. Because it's narrow, its guardrails are easy to explain. A dot is much broader, so its guardrails need more serious design.

One local note. The ways to talk to a dot that OpenAI lists today are ChatGPT, Slack, Teams, and voice calls, with text messaging coming. OpenAI's admin page says messaging through iMessage, RCS, or WhatsApp isn't available for Enterprise at launch. For teams that work in WhatsApp every day, that's worth noting before expecting Dots to plug straight into how your team works.

A friendly face isn't a guarantee of correct results

There's a design side too. OpenAI makes a dot feel like a coworker: you give it a name and choose its shape, color, eyes, glasses, and accessories, and it gets a handle like @yourname-dot. TechCrunch described the result as a "bubbly" dot avatar. That makes AI feel more approachable, which is good for adoption.

But the Parasuraman and Manzey review I covered in the human-in-the-loop article shows that people tend to over-trust automated systems, novices and experts alike, and that it can't be prevented through training or instructions. I think the friendlier the face, the more the guardrails behind it matter. That's my opinion, not an OpenAI claim.

What's still unclear

• The price of additional dots. OpenAI has only said the first dot is free on Pro and Business Premium, and that later you'll be able to add dots or scale each one's capacity. Pricing and usage limits haven't been announced. A Premium seat itself is, according to OpenAI, US$100 per user per month billed annually (US$125 monthly), and pricing can vary by country.
• Long-term reliability. Dots launched only a few days ago. Vellum, which reviewed the launch, wrote that demos and short early-access accounts can't establish weeks of correct, permissioned work. I agree with that caution.
• Availability in Indonesia. OpenAI excludes the European Economic Area, Switzerland, and the UK for Pro, and doesn't list Indonesia as an exclusion. But the rollout is gradual, so check your own account.

If you want to try it, start safe

These are my suggestions, not OpenAI's guidance:

1. Start with read-heavy, low-risk work like research, summarizing, and preparing drafts. Keep sending and changing things behind your approval.
2. Don't connect finance or customer communication right away. Add apps one at a time, and check what the dot does through Activity View first.
3. Use Custom Rules from the start, such as banning sending email, then loosen them gradually.
4. Treat a dot's memory as personal notes, not company knowledge.
5. In a company, keep control of access. If you're on Enterprise, leave dot access off until there are rules about who may use it and for what, and add dots to the checklist you run when an employee leaves.
6. Record who owns each dot and what its job is, so someone is accountable.

Wrapping up

Dots is interesting not only because it can work 24/7. What's interesting is that it forces us to answer an old question that usually gets skipped: whose task is this, what can it access, and who is accountable when it's wrong. That question applies to any agent, whether you build it yourself or buy it from a vendor, and the list of questions for that is in Before Buying AI Services for Your Company, Here Are the Questions to Ask the Vendor.

If you're weighing an AI agent for your team and want to settle its tasks, access, and oversight level first, tell me about your needs first.