OpenAI’s 24/7 AI Agent: the best personal assistant or a privacy threat?
OpenAI's dots keep working after you close the laptop.
Until this week, most of us used ChatGPT only while we were at the keyboard. You asked a question, it answered, and the work ended when you closed the tab. OpenAI's new dots break that pattern.
A dot has its own cloud computer. It can keep working after you leave, use the apps you give it access to, watch for changes, run scheduled work and come back when it needs a decision. You can go to sleep while it keeps working.
That sounds like a small product change, and I think it is a large one. OpenAI announced more than 20 things at DevDay 2026, including GPT-6.1 Sol, ChatGPT Space, Pages, new plugin automations, team tasks and a new Pro 500 tier. Dots are the announcement I would watch, because an agent that keeps working changes who, or what, is doing the work.
OpenAI has introduced three things that are easy to mix up. Dots are the agents, ChatGPT Space (singular) is the shared workspace where people and AI work together, and Pages are the living documents inside it. Think of them as the worker, the office and the document.
What a dot actually does
A dot is an always-on AI agent powered by GPT-6 Astra. It gets its own cloud computer and browser, and you can connect it to apps, and OpenAI says its plugins reach more than 4,000 of them. Give it several pieces of work and it can continue across them without you supervising every step, and it can message you when it gets stuck or needs a decision. That last capability matters most, because an autonomous agent is only useful if it knows when to hand a decision back to you.
A normal ChatGPT conversation waits for you. A dot can continue a project overnight, run recurring work, check connected information and return with progress.
For Caribbean companies dealing with North America, Europe and suppliers in several time zones, your working day no longer has to be the AI's working day.
Dots operate through ChatGPT, Slack and Microsoft Teams, and after setting one up on desktop you can talk to it in the ChatGPT mobile app. Texting is a limited beta for Pro users in the US, so Caribbean users should not count on it, and a dot cannot call you at launch.
Every extra application people must remember to open is another place a workflow dies.
Calendar, email, project management, files and business applications. The value of an agent grows when it can see the real business.
The damage when something goes wrong grows with it, which is why access is the first thing to govern.
OpenAI's event-driven plugin automations start work when something happens: a new task on a project board, a new email, a change in a connected system. ChatGPT can read the attached material and start preparing a plan before anyone types a prompt.
We have spent years learning how to ask AI questions, and these systems can begin without one.
OpenAI calls this proactive research. When you are not working with it, a dot uses read-only tools on your connected apps to look for things worth your attention, and OpenAI says those tools cannot send messages, change app content or control a browser or computer.
"Check this invoice" and "keep looking through the company until you find something interesting" are instructions with very different boundaries.
A dot builds context over time, including from connected apps, so you should not have to explain the company from scratch every Monday morning.
OpenAI says disconnecting an app leaves what the dot already obtained in place, and only deleting the dot removes it.
The part I like more than I expected: Pages
We have also been testing Pages, and they may turn out to be the quieter but more useful part of the release. The old loop was to ask ChatGPT for something, copy it into a document, edit it, return to ChatGPT, explain the changes and copy the new version back. On a page the work stays in one place: text, charts, research and images appear together, people can edit, the AI can edit, and a dot can be told to keep parts of the page current from connected sources. That removes a great deal of copying and pasting.
The early verdict is favourable. Drafting inside a page and asking for a section to be revised in place is faster than the old loop, and charts and images generated straight into the document remove a step that used to need a second tool.
There is still friction. Some work pulls you back toward the chat interface when you would rather stay inside the document, and anyone who writes for a living notices that interruption within an hour.
Connecting Slack is the workaround we are using, because updates reach the team where it already works. Collaborative slides were announced too, but OpenAI says they arrive in the coming weeks, so we cannot test them yet. We are continuing to test specific use cases and the safety questions covered below, and the direction is already clear: the document is becoming active.
- The living client brief. Put the scope, budget, constraints, decisions and outstanding questions in one page. Let the team work from it and let the dot update what changes, which ends the
FINAL_v7_REALLYFINAL.docxsitting beside three other finals. - The Monday report nobody enjoys producing. A sales update, project status, management report or board summary. Connect the approved sources, schedule the collection, and let the first draft exist before the person responsible reaches the office. The human job becomes checking the numbers, interpreting them and deciding what matters, instead of hunting through five systems for Tuesday's figure.
- Meetings that do not evaporate. A surprising amount of organizational memory disappears between "we agreed to do that" and "who was supposed to do that?" OpenAI's Meetings plugin, in beta on macOS for Pro and Business users, produces notes and action items in Space that can feed the project plan and the follow-up. That is mundane, and mundane AI is where a lot of the money will be made.
From a birthday party to a ministry: six real uses
The enterprise pitch makes dots sound like software for large companies, and the ordinary uses are broader. Each of the six below is a standing job, something done repeatedly that could keep moving when nobody is at the keyboard.
A parent planning a birthday party can keep the guest list, budget, cake, venue, shopping and reminders together on one page, with a countdown of reminders working back from the date. Share the page with the other parent so nobody buys two cakes.
Overnight, the agent reads what came in, drafts the replies that are routine, and leaves a short list of what actually needs you. The proposal stays on a page with its pricing table and chart generated in place, so a colleague gets a link rather than a fourth attachment.
A small business can have a dot watch stock against sales and flag the items that need reordering, keep a supplier page with prices and lead times current, and draft the week's posts from what actually sold. This is work a firm of eight people has nobody to do.
A government team can maintain a page per policy or project holding the consultation notes, the draft and the comments, and compile the recurring status report from approved sources each week. The record of who changed what stays beside the document rather than in an inbox.
Scheme of work, lesson plans and rubrics on shared pages a department edits together, with diagrams generated in place. A standing job to draft differentiated versions of a worksheet for different reading levels.
A page for the fundraiser with the roster, the budget and supplier contacts, and an agent that chases the outstanding confirmations. Volunteers read it on a phone; the organizer edits on a laptop. The information stops disappearing into one person's WhatsApp history.
Each diagram shows how the standing job is set up, and the items and proportions in them are illustrative. Dots are not yet available to users under 18, so every example assumes an adult account.
What job do we repeatedly do that could keep moving when nobody is at the keyboard? Then the more important question: where must it stop and ask us?
The uncomfortable part
I like this direction, and I am more cautious about it than I was about chatbots. A chatbot with bad judgment gives you a bad answer. An agent with bad judgment and permissions can do something, which puts it in a different risk class.
Days before this launch, on 25 September, OpenAI disclosed that agents in its research environment had transmitted user-provided images from training and evaluation data to third-party image-hosting sites as unlisted links, and it identified 53 instances (BleepingComputer's report). Those were not 53 newly released dots going rogue in customer accounts, and OpenAI says the cases occurred before its safeguards were in place. The episode still matters, because it shows why autonomy, access and external tools need different treatment from a model answering inside a chat window.
OpenAI itself says dots can make mistakes and that consequential work should be reviewed, and I agree.
- What can it do without asking? Write it down instead of leaving it as a philosophical question. OpenAI's custom rules let a dot take an action without asking, take it only if you pre-approved it, ask before taking it, or hand it off to you, so use them.
- Which information can it see? Connecting the company drive, Slack and email because "AI needs context" does not amount to a data strategy. Give it what the job requires, and leave access to your local computer, which OpenAI starts turned off, that way until a job needs it.
- What is it allowed to remember? A dot forms memories from connected apps, and disconnecting an app leaves what it already obtained in place until the dot itself is deleted. Offboarding an AI agent now deserves some of the thought we give to offboarding a human employee.
- Who reads the activity log? OpenAI provides an activity view of what the dot is doing, including background work, and someone should own reviewing it. A log that nobody reads records history and provides no governance.
- Where can the data go? For Jamaican organizations this is a legal question as much as an IT one. The Data Protection Act places obligations on data controllers around lawful processing, security, processors, data minimization and transfers outside Jamaica. If a dot reads customer records, moves information between systems or acts through external services, your Data Protection Officer belongs in the conversation before deployment.
- What happens when it is wrong? Every standing AI job needs a failure path: who is notified, whether the action can be reversed, what must always require approval, and how quickly access can be revoked. OpenAI lets you pause a dot and reset it, and reset deletes its conversations, saved memories and scheduled tasks. "Human in the loop" means very little until someone decides which loop, which human and when.
When your assistant talks to my assistant
These agents will increasingly interact with other agents. OpenAI already talks about teams of dots working together on behalf of one user, and the logical next stage is agents representing different people and organizations dealing with one another. That raises a social question: what happens when your assistant talks to my assistant more often than you talk to me? It raises a workplace one too: what happens to young workers entering organizations where many of the small tasks that once taught people how the organization worked are now handled by agents? Neither question calls for panic, and both deserve attention now, while organizations are still deciding how to use these agents.
Who can get dots
At launch, dots are rolling out to ChatGPT Pro users in markets outside the European Economic Area, Switzerland and the UK. OpenAI has not published a country-by-country list, so a Caribbean subscriber should confirm access rather than assume it. Business Premium users can access dots across all supported ChatGPT regions. Enterprise users, including eligible Edu and Healthcare workspaces, can try the beta when an administrator enables it, and it starts off by default.
Rollout is gradual, so two people on the same type of plan may not receive access on the same day. Your first dot is included on eligible plans, and for the next month dots usage will not count toward plan allowances, with usage terms to follow. You create a dot in the ChatGPT desktop app or on desktop web, and once it exists you can talk to it in the mobile app; you cannot create one on mobile. Dots are not yet available to users under 18. Space and Pages have broader availability across Pro, Business and Enterprise on desktop and web, with mobile reading and sharing now and mobile creation still to come.
The larger DevDay announcement
Dots got my attention, but they were part of a much larger announcement. OpenAI also released GPT-6.1 Sol, expanded cloud Codex, introduced event-driven plugin automations, team tasks, ChatGPT Space and Pages, announced collaborative slides, and launched a higher-usage Pro 500 tier. The individual features matter and the pattern matters more: ChatGPT is becoming less a website where you ask an intelligent model questions and more an operating layer for work. That changes what companies should prepare for. Prompt training alone will not be enough, because organizations will need to know how to design agent jobs, permissions, escalation rules, data boundaries, monitoring and human review, which is a different skill set.
My take
For years the AI conversation has been dominated by intelligence: which model is smarter, which benchmark went up, which one writes better. Dots make me more interested in another variable: the number of hours useful intelligence can keep working without me.
Frequently asked questions
A Caribbean company with eight people has always had more work than hands. Now some of that work can continue after everybody goes home, which is a very good reason to decide what the machine is allowed to do before giving it the keys.Adrian Dunkley · President, Caribbean AI Association · 29 September 2026
When the Agent Goes Rogue: Five Incidents, Explained Plainly
What happens when a system with permissions keeps acting, and the controls to set before switching agents on.
Read the piece →The Wrapper Economy: A Caribbean Buyer's Guide to Warning Signs and Risks
How to tell a real product from an interface over somebody else's model, and what to ask a vendor.
Read the guide →Jamaica Trained an AI Model to Catch the Bias in Everyone Else's
Why a small model that runs on local hardware answers a different question from the frontier systems.
Read the piece →The Caribbean AI Newsletter is the leading daily source for artificial intelligence news, analysis, and practical insight for the Caribbean. It covers policy, workforce, education, cybersecurity, climate resilience, governance, and the tools Caribbean organizations actually use.
