OpenAI's Dev Day was covered everywhere as a developer event — "OpenAI COOKED," "massive announcements," the usual launch-day fever. But the headline product, ChatGPT Dots, isn't really a developer product. It's an always-on agent that lives in the apps a business already runs. That makes it a buying question, and almost nobody is answering it for the owner who doesn't write code.
Here's that briefing: what shipped, what you can use without a developer, what to wait on, and the rules that decide whether your dot is an asset or a liability.
What a dot is, in plain English
A dot is a named, always-on agent inside ChatGPT. Three things separate it from the chat window you already know:
- It keeps working when you close your laptop. Each dot gets its own cloud computer and browser, so it can hold an ongoing goal and make progress between conversations. It messages you in Slack or Microsoft Teams when it needs a decision.
- It connects to your tools. 4,000+ apps through the plugin catalog, plus Slack and Teams.
- It remembers. It inherits your existing ChatGPT memories and builds new ones from the apps you connect — which is why the order you connect things matters (more on that below).
Under the hood it runs GPT-6 Astra. When it's idle, it looks for ways to help using read-only tools. The pitch, in one line: an assistant who watches your shop overnight and only knocks when something needs you.
Can you actually get one today?
Check this list before you plan anything:
- ChatGPT Pro and Business Premium: the first dot is included. No separate purchase.
- Free, Plus, and lower tiers: not eligible. If you're on the entry-level plan, there's nothing to evaluate yet.
- Enterprise, Edu, and Healthcare: beta, off by default. A workspace admin has to switch it on.
- Setup is desktop-only. You can message your dot from mobile afterward, but you can't create or configure one from a phone.
- The rollout is gradual. Eligible doesn't mean "visible today" — some accounts get features days later.
- Pro users in the EEA, Switzerland, or the UK: not yet. Business Premium workspaces in those regions are fine.
Three buckets: use Monday, wait, or developer-only
Use Monday, no developer required:
- Give it one boring, ongoing watch job: scan your reviews daily and draft replies for approval, watch a competitor's pricing page, or summarize tomorrow's calendar every evening.
- Set the permission rules (next section). It's plain-language configuration, not code.
- Let it draft — reports, recaps, replies — while you keep the send button.
Wait on these:
- Month-two economics. Chatting with your dot doesn't count against your usage limits, and background work isn't counted during the first month. After that, per-tier terms are unpublished. Don't build a workflow around a price nobody has announced.
- Texting. A US-only beta for Pro users.
- Teams of dots. Planned, not shipped. One primary dot per person today.
Developer-only for now:
- The Codex hand-off (the dot routing coding work to OpenAI's coding agent).
- "Specialist dots" with their own identities and credentials — enterprise pilots only.
The rules matter more than the agent
Every action a dot can take maps to one of four permission levels. Set these before you connect anything:
- Take action without asking — save this for drafting, summarizing, and internal notes.
- Take action if pre-approved — narrower than it sounds: it means you asked for that exact action in your request, not that you approved a category in advance.
- Ask before taking action — anything customer-facing: emails to customers, calendar changes involving outsiders, replies on public pages.
- Hand off to you — payments, contracts, anything legal or irreversible. You do these, always.
There are also brakes you can't switch off: proactive background work is read-only, an auto-review system screens actions that could affect your accounts before they run, and password changes, data deletions, and software installs always come back to you.
The advice: default to "ask before taking action" on anything that leaves your company, and let the dot earn autonomy on internal drafts and research. An agent's first month should be boring on purpose.
The honest catches
- Nothing is independently verified. No benchmarks, no third-party audits. OpenAI's own documentation says a dot can make mistakes, including while following your rules. Treat early testimonials as anecdotes, not results.
- The failure class is real. Four days before launch, OpenAI disclosed that a separate system — internal research agents — posted 53 users' images to public hosts. Different product, but it's exactly the kind of incident to keep in mind before you hand an always-on agent real credentials.
- Memory is one-way. You can't delete a dot's memories without deleting the dot, and disconnecting an app doesn't erase what it already learned from it. Connect low-stakes apps first. Sensitive inboxes come last — or never.
The one-line strategy
Dots is the third "agent with its own cloud computer" shipped in seven weeks — Meta's Muse on September 8, Grok Bot's beta in August, Dots on September 29. The platforms are in a land grab for your workflows.
So don't pick a winner. Keep your data exportable and your workflows written down somewhere the agent doesn't own, so that when the next one ships — and it will — you can add it or move to it without rebuilding. The businesses that benefit from this cycle aren't the ones that bet on the right agent. They're the ones that stayed cheap to switch.
A first-week plan
- Check eligibility on desktop: your plan, your region, and whether Dots has reached your account yet.
- Connect low-stakes apps first — calendar, review feeds, a folder of internal documents. Not the customer inbox. Not payments.
- Write the rules before the first real task: ask-before on external sends, hand-off on money and legal.
- Give it one boring ongoing goal. Example: "Watch our reviews daily. Draft replies for my approval. Flag anything three stars or below immediately."
- Spot-check every morning for the first week: what it did overnight, what's queued for your approval, what it learned.
- Before month two, read the published usage terms — then decide what else it's allowed to touch.
The demo that matters isn't the launch keynote. It's a dot that flagged a bad review before you'd had coffee, drafted the reply, and waited — because you told it to.
Want the permission-ladder templates and first-week checklists for your first AI agent? The AI Automation Starter Kit includes rule-writing frameworks, approval gates, and the guardrail checklists that keep an always-on agent from becoming a liability.
Top comments (0)