What Can OpenAI Dots Do? The Problems They Actually Solve

Every chat assistant has the same flaw: it forgets the job the moment you stop typing. You ask ChatGPT to find a missed invoice, it tells you how, and the invoice stays unsent. So what can OpenAI Dots do about that? OpenAI's answer, in the launch post, is an agent that holds the goal, owns a computer and keeps going. Below are the problems that design solves, matched to the examples OpenAI and early testers have published, and the ones it still leaves to you.

What can OpenAI Dots do that ChatGPT cannot?

The problem	ChatGPT today	A dot
Work stops when you close the tab	Each chat ends when you do	Runs around the clock on its own cloud computer
The task lives in your apps	Reads connectors on request	Signs into 4,000+ apps through plugins and acts in them
Something happened while you were away	Nothing, until you ask	Background research with read-only tools, then a suggestion
A job needs a decision from you	Waits in a thread	Pauses, asks in Slack, Teams, voice or ChatGPT, resumes
Your standards are never remembered	Memory across chats	Notes to itself and learns from feedback on each task

The rest of this guide takes each row in turn.

Problem 1: the work ends when the conversation does

A dot works "24/7", per OpenAI, toward a goal you set once. Between your messages it runs what OpenAI calls proactive research: scanning the sources you connected for things worth doing, then raising them with you. The safety post fences that mode tightly. Research runs on read-only tools, cannot message anyone, cannot change anything in a connected app and cannot touch a browser or desktop.

The reported two-day test from early October shows the shape of it. The tester asked for couch options, got a three-page packet of four, rejected them, then specified colours and a pull-out bed and asked for ten ranked by a points rubric. The second pass landed closer, and the dot was then set to watch the product pages for sales. That watch is the part no chat assistant does: nobody had to come back and ask again.

Problem 2: the invoice that never got sent

OpenAI's own example is small and telling. An early tester's dot noticed an invoice to a publication had never gone out, prepared it, and sent it after the tester approved. Three things had to be true for that to work. The dot could read the accounting or email source. It could draft inside the app rather than paste text back. And the final send was gated by an approval.

That gate is the design, not a limitation. Purchases on a saved card, permanent deletions and installing unfamiliar software all wait for you. Moving money between accounts or changing a password is handed back to you to finish yourself, with the dot allowed to do the surrounding steps. Health information can only be sent to a named recipient you specify.

Problem 3: the bug report nobody picks up

"A bug appears in Slack, and dots immediately start investigating," the announcement says. Because dots sit inside Slack and Microsoft Teams, the trigger is the channel message, not a prompt. A dot with access to the repository and the error logs can reproduce the issue, narrow it down and post what it found before an engineer has read the thread.

The second engineering example is bigger: "A new design arrives, and dots turn it into a working app while the team focuses on customer feedback." That depends on Codex, and work run through Codex or ChatGPT Work counts against those allowances even though plain dot conversations do not. Teams using ChatGPT for code should budget for that.

Problem 4: the project that drifts between meetings

"A planning cycle gets underway, and dots keep everyone moving in sync toward the deadline." OpenAI gives no mechanics, but the pieces are documented: a dot can read the tracker and the calendar, carry context across every channel and nudge people where they already are. Activity View shows what it is doing, including background work, and lets you redirect it. Custom Rules let you say, for example, that it never emails a client without approval.

The system-card figures are the caution here. In OpenAI's chained-task test, reported by The Rundown, moderate scope-violation flags rose from 8.6% to 19.7% as intervening tasks went from five to ten, with no severe breach. A project dot handles many small tasks in sequence, which is exactly where that drift shows up, so review its work in the Activity View rather than trusting the summary.

Problem 5: back-office queues no one wants to own

For companies, OpenAI is piloting specialist dots with their own identity, credentials and system access. The pilot areas are procurement, invoice processing, email marketing, customer support and commercial contracting. Each one is a queue of similar, rules-bound tasks, which is the kind of work an always-on agent can take without a human assigning each item.

Governance is the sticking point for IT, and OpenAI is addressing it through Microsoft rather than alone: specialist dots are being integrated with Agent 365, so a business can manage them through the same identity and audit tools it uses for other agents. Enterprise, Edu and Healthcare workspaces get the beta switched off by default until an admin enables it.

What dots do not solve yet

• CAPTCHAs and bot walls. In the same early-October test, a puzzle CAPTCHA on a subscription site stopped the dot from cancelling, and it asked the tester to take over.
• Editable memory. You cannot view, correct or delete individual dot memories, per OpenAI's help centre. Deleting the dot is the only reset, and disconnecting an app does not remove what it already learned.
• Prompt injection. OpenAI says its layered checks reduce the risk from malicious instructions hidden in web pages and emails, and does not claim to remove it.
• Everything behind a paywall. The first dot needs ChatGPT Pro or Business Premium, with Pro unavailable in the EEA, UK and Switzerland at launch. For the plan details see our OpenAI Dots explainer.

Who should turn one on

A dot earns its place when you have recurring work across apps you can let it into, with a clear rule for what needs your sign-off. Solo operators with invoicing and inbox chores, engineering teams in Slack and back-office teams with queues fit. If your plan does not include one, Lindy builds email and CRM agents on a credit plan, and Zapier agents trigger on events across its app catalogue; our ranking of AI automation tools compares them.

FAQ

Can OpenAI Dots send emails for me?

Yes, with an approval gate. A dot can draft and send from connected email, and Custom Rules can require your sign-off on every send or block sending entirely. Background research mode can never send a message.

Can a dot buy things?

It can shop, compare and prepare a purchase, but paying with a saved card requires your approval each time, per OpenAI's safety post. Transfers between financial accounts are handed back to you to complete.

Does a dot work while I sleep?

Yes. It runs on its own cloud computer, not your laptop, and carries on with set goals and read-only research overnight. Anything that needs an action outside your rules waits for your approval in the morning.

Can OpenAI Dots use my laptop?

Only if you grant that permission, which is off by default. Everything else runs in the dot's sandboxed cloud workspace.

What can OpenAI Dots do with company data?

In Business, Enterprise and Edu workspaces OpenAI does not train on your content by default. Human review can still happen in limited safety cases, and the dot keeps what it has learned until it is deleted.