Always-on AI agents arrived: what an agent that never logs off means for your business

Q: What did OpenAI announce with dots?

On September 29, 2026 OpenAI introduced dots, always-on agents powered by its GPT-6 Astra model. Each dot has its own cloud computer with a web browser, connects to more than 4,000 apps, learns preferences over time and can be reached through ChatGPT, Slack or Microsoft Teams. Dots follow ChatGPT's permissions and support custom rules for which tasks require approval. They are rolling out to Pro, Business Premium and Enterprise plans in select markets, with one dot included at no extra cost.

Q: Should a small business turn on an always-on agent?

Yes, but for one job, not for everything. The background pattern is exactly where a smaller business gains the most, because the work that piles up overnight - inbound orders, supplier emails, overdue invoices - has no one watching it. Start with a single standing goal on one process, give the agent read access to the systems that process touches, have it draft rather than send, and review its output daily for the first few weeks.

Q: What are the risks of an agent that works 24/7?

The main risk is unattended action at scale. An agent connected to thousands of apps with a broad goal can take many small wrong steps before anyone looks. The controls are the same ones that make any agent safe, applied more strictly because nobody is watching live: a narrow goal, its own identity with scoped permissions, read-only by default, human approval on anything customer-facing or financial, a full activity log and an off switch.

An always-on AI agent is an agent that keeps working toward a standing goal after you close the chat. It does not wait for a prompt. It runs in the background on its own computing environment, watches the systems it is connected to, remembers what happened yesterday and acts or reports when something relevant changes. Until now most people met AI as a conversation: you ask, it answers, the session ends. On September 29, 2026 OpenAI made the other model a mainstream product, with agents that are meant to work on your goals around the clock. For a traditional or small business, the interesting part is not the launch itself. It is that the background work piling up in every company overnight finally has a realistic owner, and that the rules for letting software act unattended now matter more than the model.

Key takeaways. On September 29, 2026 OpenAI introduced dots: always-on agents with their own cloud computer, access to more than 4,000 apps, and memory that improves with feedback, reachable through ChatGPT, Slack or Teams. The shift is from "answer when asked" to "work on a goal you set once". For a smaller business that is a real opportunity, because the costly work is often the work nobody watches between 6pm and 9am. The catch is that unattended action multiplies mistakes as fast as it multiplies output. The winning pattern is one standing goal, on one process, with its own scoped identity, drafts instead of sends, and a human approving anything customer-facing or financial.

What OpenAI actually shipped

Here are the facts, without the hype. OpenAI introduced dots at its DevDay event on September 29, 2026. Each dot runs on the GPT-6 Astra model and has its own cloud computer with a web browser, so it keeps working after you close your laptop. It connects to more than 4,000 apps, carries context across channels and, in OpenAI's words, can work toward your goals 24/7. Moreover, it learns your preferences from feedback over time. Dots are rolling out to Pro, Business Premium and Enterprise plans in select markets, and according to MacRumors they follow ChatGPT's permissions and support custom rules for which tasks require approval. That last detail is the most important line in the announcement. The vendor itself is telling you that an agent working unsupervised needs an explicit line between what it may do alone and what it must ask first.

Why the background is where a small business wins

Most of the expensive work in a smaller company is not hard. It is unwatched. An order email arrives at 11pm and sits until someone opens the inbox. A supplier changes a delivery date and nobody notices until the truck is late. An invoice goes overdue for three weeks because chasing it is nobody's first task. Large enterprises throw headcount and shift rotas at this; a 40-person distributor or a family manufacturer cannot. That is precisely why the always-on pattern fits the smaller business better than the enterprise. In fact, the biggest gap is not adoption but integration. A Goldman Sachs 10,000 Small Businesses survey of 1,256 program participants found 76% already use AI, yet only 14% are fully integrating it into core operations. A background agent on one real process is how you cross that line, as our small-business playbook describes.

The catch: unattended means unwatched

There is a trade-off worth stating plainly. A chatbot that gives a bad answer wastes a minute, because a person reads the answer before doing anything. An always-on agent that takes a bad step at 3am may take forty more before anyone looks. Consequently, the risk is not that the model is worse, but that the human checkpoint moved from every step to the morning review. A goal like "keep my projects moving" across thousands of connected apps is far too broad to audit. The same goal narrowed to "flag every inbound order that does not match the price list, and draft the reply" is easy to check, easy to measure and easy to stop. This is why the controls we use for any agent apply more strictly here: its own identity with scoped permissions, read-only by default, approval on writes, and a test set of real past cases before you trust it overnight.

Where your data lives still matters

Furthermore, an always-on agent runs somewhere, and in the consumer-style products that somewhere is the vendor's cloud. For a lot of everyday work that is fine. However, a standing agent that reads your customer list, your pricing and your supplier contracts every hour is a different exposure than a one-off question in a chat window. Before you switch on any background agent, decide which systems it may read, under which credentials, and whether that data is allowed to leave your perimeter at all. Often the right answer for a traditional business is the same pattern we described for self-hosted agents: run the background job inside your own environment, on your own keys, against your existing systems. You get the 24/7 benefit without negotiating a new data question with every vendor that ships an agent.

What to do this month

Write down one standing goal. Pick a single process where work waits overnight, such as order intake, overdue invoices or supplier changes, and phrase the goal narrowly enough that you can check the agent's output in five minutes. Our implementation guide shows how to scope it.
Give the agent its own scoped access. A dedicated account, read-only on the systems that process touches, nothing else.
Draft, do not send. Have the agent prepare replies, flags and updates, and keep a human approving anything customer-facing or financial. Our meeting-prep agent works exactly this way.
Review a morning digest. For the first few weeks, read a daily summary of what the agent did, correct it, and measure hours saved against a number you wrote down beforehand.
Expand only after it earns trust. Loosen one approval or add one process at a time, the same sequence we use for every client automation.

Frequently asked questions

What is an always-on AI agent?

An agent that keeps working toward a standing goal after you close the chat. It runs in the background, watches connected systems, carries context between days and acts or reports when something relevant happens, instead of waiting for a prompt.

What did OpenAI announce with dots?

On September 29, 2026 OpenAI introduced dots: always-on agents on GPT-6 Astra, each with its own cloud computer, access to more than 4,000 apps and reachable through ChatGPT, Slack or Teams. They support custom rules for which tasks need approval and are rolling out to Pro, Business Premium and Enterprise plans in select markets.

Should a small business turn on an always-on agent?

Yes, for one job. Pick a single process where work waits overnight, give the agent read access to that process only, have it draft rather than send, and review its output daily before you widen its scope.

What are the risks of an agent that works 24/7?

Unattended mistakes at scale. The fix is a narrow goal, the agent's own scoped identity, read-only by default, human approval on customer-facing or financial actions, a full activity log and an off switch.