--- title: "OpenAI Dots Proves Always-On Agents Are Here. For EU Companies, the Real Question Is Where They Live." canonical: "https://agentbarn.dev/blog/openai-dots-always-on-agents" author: Gaurav description: "OpenAI dots makes always-on agents a category. For EU companies under NIS2 and the AI Act, the real question is where those agents run and who holds the keys." categories: [Blog, AI agents, EU AI Act, NIS2, Self-hosting] --- # OpenAI Dots Proves Always-On Agents Are Here. For EU Companies, the Real Question Is Where They Live. On September 29, [OpenAI launched dots](https://techcrunch.com/2026/09/29/openai-launches-dots-its-bubbly-agentic-avatar/). A dot is an agent that doesn't stop when you close the tab. It has its own computer, it [connects to over 4,000 apps](https://www.pymnts.com/news/artificial-intelligence/2026/openai-launches-dots-to-capture-ai-agent-market/), and it keeps working while you sleep. You talk to it in Slack or Teams, like a colleague. Most people will read this as a product launch. I think it's something more useful than that. It's the moment always-on agents stopped being a thing early adopters argued about and became a category. Meta's Muse [passed 3 million downloads](https://www.vellum.ai/blog/official-openai-dots-breakdown) in the three weeks before dots arrived. When the two biggest consumer AI companies ship the same idea within a month of each other, the idea is no longer the interesting part. So what is the interesting part? For a company of 100 people in Europe, I think it's a question almost nobody is asking yet: where should these things live? ## An agent has keys Here's the thing about agents that's easy to miss. A chatbot only knows what you paste into it. An agent has keys. It has to. An agent that can't log into your CRM, read your inbox, or update your tickets isn't an agent. It's a chatbot with a scheduler. The whole value of an always-on agent comes from standing access to your business, used without you watching. That's not a flaw in the design. That is the design. Which means the question "should we use AI agents?" is the wrong question. Of course you should, eventually. The right question is the one you'd ask about anyone you gave a master key to: where do they work, who can see what they did, and what happens if you want the key back? With dots, the answers are clear enough. Each dot runs on [its own cloud computer](https://thenextweb.com/news/openai-dots-always-on-ai-agents-cloud-computers-devday), which OpenAI hosts. Your credentials and your data go there. You'll notice Europe got a slightly odd version of the launch: individual Pro users in the EEA, Switzerland, and the UK were left out, but [Business Premium customers get dots everywhere ChatGPT is supported](https://www.pymnts.com/news/artificial-intelligence/2026/openai-launches-dots-to-capture-ai-agent-market/). So your company can buy it today. The question is whether that's where you want your keys. ## The rules that decide things in Europe Now the boring part, which in Europe tends to be the part that decides things. ### The EU AI Act Start with the EU AI Act, because everyone will. The strictest rules were supposed to hit in August. They didn't. The [Digital Omnibus](https://usercentrics.com/knowledge-hub/eu-ai-act-high-risk-delay-article-50-transparency-consent/) pushed the high-risk deadline to December 2, 2027. A lot of people heard "delay" and relaxed. I think that's a mistake. The deadline moved because [standards and regulators weren't ready](https://www.joneswalker.com/en/insights/blogs/ai-law-blog/yes-august-2-still-matters-the-eu-approved-a-high-risk-ai-delay-but-most-trans.html?id=102nbon), not because anyone changed their mind about the rules. To be fair, most of what agents do isn't high-risk. An agent that writes your Friday report is not what Brussels is worried about. But companies don't decide to build high-risk systems. They drift into them. The agent that summarises stand-ups starts noticing who's always behind. The inbox agent starts sorting job applications. Each step is small and helpful. Then one day you're doing performance monitoring and recruitment screening, which are exactly the areas the Act lists. At that point you need logs you can keep, and people who can oversee what the system does. It's much easier to have built that in than to bolt it on. ### NIS2 The rule more likely to affect you today is NIS2. It applies to [companies with 50 or more employees in 18 sectors](https://www.cm-alliance.com/cybersecurity-blog/are-you-in-scope-of-nis2-the-complete-entity-classification-checklist), including a lot of manufacturing, food, chemicals, and logistics. In Germany alone it pulled [around 29,500 companies](https://de.secfix.com/post/germanys-new-nis2-cybersecurity-law) under supervision, most of them mid-sized and new to this. NIS2 wants you to secure your supply chain and report incidents fast, with [a first warning within 24 hours](https://www.advisori.de/en/blog/nis2-mid-sized-companies-practical-guide-2026). An agent with keys to your ERP is part of your attack surface. Its vendor is part of your supply chain. When something goes wrong, "we've opened a ticket with the vendor" is a bad answer. ### Works councils And if you're in Germany or Austria, your works council gets a say too. Technology that can monitor employees usually needs their agreement, and an agent that reads Slack, email, and Jira can obviously monitor employees. They'll ask what it reads, who sees the output, and how long records are kept. It's a lot easier to have that conversation about a system you can open up and show them than about a black box in someone else's data centre. ## Safeguards in someone else's house We should be fair to OpenAI here, because they clearly thought about this. Dots have [approval gates](https://www.theneurondaily.com/p/openai-launched-dots-20-more-tools) for consequential actions. Saved passwords are used [without being shown to the model](https://thenextweb.com/news/openai-dots-always-on-ai-agents-cloud-computers-devday). They're [working with Microsoft](https://www.medianama.com/2026/10/223-openai-launches-dots-devday-2026/) on governance controls for companies. These are real safeguards made by smart people. But they're safeguards in someone else's house. The vendor decides the controls, the log format, the retention, and the price. OpenAI says it will [publish usage terms per plan later](https://www.pymnts.com/news/artificial-intelligence/2026/openai-launches-dots-to-capture-ai-agent-market/). That's normal for a launch. It's also a reminder of who holds the pen. So here's my honest view. If you're a 15-person agency with no regulated data, use dots. It takes minutes, and the trade-offs won't hurt you. Dots is also fine for personal work, like research and drafting, at any size. The calculation changes when agents start acting inside your core systems: customer data, people decisions, anything NIS2 cares about. There, control matters more than convenience, and control mostly comes down to where the thing runs. ## Where Agent Barn fits We should say now that we're not neutral. We build [Agent Barn](https://agentbarn.dev/), which is the other answer to the where-do-they-live question. So discount what follows accordingly. But I'll tell you why I think it's the right answer for most mid-sized European companies, and you can judge the argument on its merits. Dots and Agent Barn are built on the same insight. Agents should live in the chat tools your team already uses. They should run on schedules and react to events. Each one should have its own identity and its own scoped access. When OpenAI shipped dots, our first reaction wasn't worry. It was relief. They spent a keynote explaining our idea to the world. Where we differ is the one thing that matters most in Europe. Agent Barn runs on your own Kubernetes cluster, so your credentials and data never leave infrastructure you control. You pick the model, too. It routes through LiteLLM, so you can use a frontier API, a model hosted in the EU, or an open-weight model on your own hardware. It's [open source](https://github.com/aai-labs/agent-barn). That sounds like a developer preference, but it's actually a business one. Your auditor can read it. Your works council can read it. And if we ever stop being worth paying, you keep everything: your agents, their configuration, their history. A vendor can't change the terms on software you already own. And every conversation, tool call, and cost is logged in your own database. I think this will turn out to be the most underrated feature in the category. Nobody gets excited about logs. But when a regulator, a customer, or your own board asks what an agent did last March, the company that can answer in five minutes from its own records is in a completely different position from the one waiting on a vendor export. The obvious objection is that a 100-person company doesn't want to run Kubernetes. That's fair. Most don't. That's why we start every deployment as a pilot, set it up with you, and hand it over with support. You need someone to own it afterwards, but you don't need a platform team. ## What to do this quarter So what should you actually do? Don't write an AI strategy. Pick one job that eats hours every week and needs a lot of context: quoting, order status, intake, the Friday report. Put one supervised agent on it. Measure what it saves. Keep a record of everything it does. Whatever vendor you talk to, including us, ask them the same three questions you'd ask about anyone you hand a master key to. Where does it run? Who holds the keys? Where are the logs, and can we keep them? If the answers are vague, that tells you something. If you do this now, you get two things for the price of one. You get an agent doing real work this quarter. And you get the habits, the logging and the oversight, that the regulations will expect from you anyway. The companies that start in late 2027 will be doing the same work under a deadline. Always-on agents are coming to your company whether you plan for it or not. The only real choice is where they live. I'd make that choice on purpose. If you have a workflow you think is worth automating, [tell us about it](https://agentbarn.dev/#intake). _This is general information, not legal advice. Check your obligations under the AI Act, NIS2, and labour law with qualified counsel._ ## Sources - [TechCrunch: OpenAI launches Dots](https://techcrunch.com/2026/09/29/openai-launches-dots-its-bubbly-agentic-avatar/) - [PYMNTS: OpenAI Launches Dots to Capture AI Agent Market](https://www.pymnts.com/news/artificial-intelligence/2026/openai-launches-dots-to-capture-ai-agent-market/) - [The Next Web: OpenAI launches dots](https://thenextweb.com/news/openai-dots-always-on-ai-agents-cloud-computers-devday) - [MediaNama: OpenAI launches Dots, plus other DevDay updates](https://www.medianama.com/2026/10/223-openai-launches-dots-devday-2026/) - [The Neuron: OpenAI launched Dots + 20 more tools](https://www.theneurondaily.com/p/openai-launched-dots-20-more-tools) - [Usercentrics: EU AI Act Digital Omnibus now in force](https://usercentrics.com/knowledge-hub/eu-ai-act-high-risk-delay-article-50-transparency-consent/) - [Gibson Dunn: EU AI Act Omnibus Agreement](https://www.gibsondunn.com/eu-ai-act-omnibus-agreement-postponed-high-risk-deadlines-and-other-key-changes/) - [Jones Walker: August 2 still matters](https://www.joneswalker.com/en/insights/blogs/ai-law-blog/yes-august-2-still-matters-the-eu-approved-a-high-risk-ai-delay-but-most-trans.html?id=102nbon) - [Agent Barn](https://agentbarn.dev/) - [Vellum: Official OpenAI Dots Breakdown](https://www.vellum.ai/blog/official-openai-dots-breakdown) - [CM Alliance: Are You in Scope of NIS2?](https://www.cm-alliance.com/cybersecurity-blog/are-you-in-scope-of-nis2-the-complete-entity-classification-checklist) - [Advisori: NIS2 for Mid-Sized Companies](https://www.advisori.de/en/blog/nis2-mid-sized-companies-practical-guide-2026) - [Secfix: Germany's NIS2 Law Is in Force](https://de.secfix.com/post/germanys-new-nis2-cybersecurity-law)