OpenAI Dots: What Always-On AI Agents Mean for Your Business

OpenAI has spent the past year promising that agents would be the next big thing, and at the end of September it finally showed what it means by that. At its DevDay conference the company introduced OpenAI Dots, always-on AI agents that keep working after you close the app. Each dot runs on GPT-6 Astra, OpenAI's current top model, and gets its own cloud computer with a browser, so it can research, draft and organise things while you are doing something else entirely.

The pitch is simple. Instead of a chatbot that answers when asked, you get a colleague that never really goes home. It remembers your preferences, connects to more than 4,000 apps, and takes instructions through Slack or Microsoft Teams as well as ChatGPT itself.

As usual with OpenAI launches, the details matter more than the demo. Dots are expensive, they are not available everywhere, and they come with real questions about memory and control. Let's sink in and see what this actually means for a business.

How OpenAI Dots Actually Work

A dot is not a chat window with a new name. When you create one, OpenAI spins up a dedicated computer for it in the cloud, a small Linux machine with its own Chrome browser, isolated from everyone else's. The dot lives there permanently. You can give it a task in ChatGPT, close your laptop, and the agent carries on browsing, reading documents and preparing drafts on its own machine.

Between tasks, dots do something more unusual: they proactively research the apps you have connected. If your dot has access to your inbox and calendar, it may go through new material while you are offline, so it has context ready when you come back. OpenAI has limited this background mode to read-only. A dot cannot send messages or change files during proactive research, and a separate auto-review system checks its actions before anything consequential, like sending an email, actually goes out.

The practical surface is wide. Dots connect to over 4,000 apps through the existing ChatGPT ecosystem, accept instructions from Slack and Microsoft Teams, handle scheduled and recurring tasks, and can even take a voice call, though only when you start it. For now you can only create them on desktop, not on mobile.

The Price of Always-On

Dots are deliberately positioned as a premium product. There is no access at all on the Free, Go or Plus plans. The entry point is ChatGPT Pro at $100 per month, which includes your first dot, or Business Premium at $125 per user per month. Enterprise, Edu and Healthcare customers get a beta that administrators have to switch on. Pricing for a second dot or for extra capacity has not been announced yet.

Sam Altman was open about the reason: these agents use a lot of compute, so they are starting out as a premium product, with mass-market versions expected later. Each dot is, after all, a full cloud computer running one of the most expensive models OpenAI has. Chatting with your dot does not count against usage limits, but the tasks it performs draw on your plan's allowance, so a busy dot can eat through a Pro subscription faster than you might expect.

There is one more catch that matters for European readers: Pro users in the EEA, Switzerland and the UK are excluded at launch. Business Premium is available in all regions, which tells you something about where OpenAI expects the real demand to come from.

What a Dot Can and Cannot Do on Its Own

OpenAI clearly learned from two years of agent experiments, because the control model is the most thought-through part of the launch. For each kind of action you can choose one of four behaviours: the dot acts freely, acts only where you pre-approved it, asks before acting, or hands the task back to you.

Some limits are fixed regardless of settings. A dot cannot change passwords, transfer money between accounts or make card purchases on merchant sites without a human in the loop, and permanently deleting data requires approval every single time. Passwords never pass through the model at all; they go through secure forms instead. Sharing health-related information requires naming the exact recipient first. And approving one message does not give the dot standing permission to keep contacting that person.

Memory, Privacy and the Prompt Injection Question

The weaker side of the launch is transparency. A dot inherits memories from your ChatGPT account and then forms its own as it works, but you cannot view, edit or delete individual memories. If something has gone wrong in there, your only option is deleting the whole dot and starting over. There are no audit log exports either, which will be a problem for any regulated business.

Prompt injection, where malicious content on a web page or in an email tricks an agent into doing something it should not, remains an open issue. OpenAI's own wording is careful: its defences help reduce the risk rather than eliminate it. Early users have already reported surprises, including one developer who found that an access token he believed was minimal actually gave his agent more permissions than intended.

Training defaults depend on the plan. Business, Enterprise and Edu data is excluded from model training by default, while personal plans follow each user's data settings. Before anyone in your company connects a dot to real systems, it is worth checking which side of that line you are on.

Where Dots Fit in a Crowded Agent Market

Dots do not arrive in an empty field. Meta launched its Muse assistant in early September with a free tier, and it quickly climbed the App Store charts as a consumer product. Microsoft has just rebuilt Copilot around its own persistent agent, Autopilot, which we covered here recently. Cognition's Devin has passed a billion dollars in annualised revenue by doing one job, coding, very well.

OpenAI is aiming Dots at professionals rather than everyday users, and the pricing says the same. There is also a separate enterprise track, specialist dots for shared company roles like support or invoice processing, which is still in pilot and requires working directly with OpenAI's engineers.

The important distinction for teams is that a dot is a personal agent. It learns one person's preferences and works under one person's account. It is not designed for shared workflows that need common audit trails and company-wide policy, and pretending otherwise will cause pain later.

Practical Takeaways for Businesses Building Digital Products

So what should a business actually do with all this? First, resist the urge to buy seats for the whole team. At $100 to $125 per person this is a tool that has to pay for itself, and the honest way to find out is a small pilot: one or two people, a handful of recurring tasks like report preparation, inbox triage or proposal upkeep, and a simple before-and-after measure of hours saved.

Second, treat permissions as the real project. The early incidents around Dots are not model failures, they are scoping failures: tokens and integrations that granted more than anyone realised. Whoever runs your pilot should create narrow credentials for the agent, separate from any human account, and review what it actually did in the first weeks.

Third, if you build digital products, think about dots as a new kind of visitor. Agents with their own browsers are now reading websites, filling forms and comparing offers on behalf of paying customers. Clean structure, working forms and machine-readable data stop being nice-to-haves; they decide whether an agent can buy from you at all.

And if you are in the UK or the EEA, the launch exclusion gives you a forced pause. Use it to sort out the permissions question and watch what early adopters elsewhere learn, rather than treating it as falling behind.

Final Notes

Dots are the clearest statement yet of where OpenAI thinks this is all going: not better chat, but persistent software colleagues that work while you sleep. Version one is pricey, desktop-bound, partially unavailable in Europe and rough around memory controls, yet the direction feels unmistakable.

Our advice is the boring kind: start small, measure honestly, and keep a human close to anything that touches money or customers. The businesses that benefit first will not be the ones with the most dots, but the ones that learn fastest what can safely be handed over. If you want help thinking through where agents fit in your own product or operations, that is exactly the kind of conversation we enjoy.