# OpenAI Dots can keep working. What can it do—and what counts against your plan? > Dots' background research is read-only, but requested tasks have separate approval rules. Here's how to decide what your agent may do and which work counts against your ChatGPT allowance. - Published: Oct 2, 2026 - Author: PiRouter Team, Engineering - Tags: agents, security, cost - Canonical: https://pirouter.ai/blog/dots-permissions-usage --- **OpenAI Dots can stay on a task after you close your laptop, but “always on” doesn't mean “approved to do anything” or “unlimited work.”** Background *proactive research* uses read-only tools on connected apps. Actions you request follow built-in rules, Custom Rules and auto-review; some, including password changes, remain yours. Conversations with a dot don't count toward ChatGPT usage limits, while Codex and ChatGPT Work tasks it starts or manages still count ([OpenAI's Dots announcement](https://openai.com/index/introducing-dots/)). Imagine a small team preparing an event launch. A dot notices an outstanding invoice, assembles the information and offers to send it. What can it read while you're away? Does “send” require your approval? If the same conversation turns into a Codex task to fix a registration page, which allowance is used? This is a hypothetical workflow, not a test we ran. OpenAI's [published example](https://openai.com/index/introducing-dots/) of an early tester's invoice is narrower: the dot prepared it and sent it **after the tester approved**. That one example helps frame the questions, but it is not a universal rule for every invoice. ![OpenAI's official Dots launch-film still presents the persistent assistant; it does not document this article's hypothetical invoice workflow.](/blog/images/dots-permissions-usage-official-intro.png "Official launch-film still, not a hands-on test or verified action outcome. Source: [OpenAI DevDay 2026 Recap](https://openai.com/index/devday-2026-recap/)") ## An assistant on a cloud computer isn't automatically on yours OpenAI introduced Dots on **September 29, 2026**. A primary dot has its own cloud computer and browser, can work with apps the user connects, and can keep several projects in progress. You can inspect that computer while it works. The user's computer and its contents stay separate by default; connecting a laptop is a further permission the user can choose to grant. The existence of the dot's browser therefore tells you nothing about access to your own browser ([Introducing dots](https://openai.com/index/introducing-dots/)). The distinction matters before the first app connection. Giving a dot access to a project board to find outstanding tasks, giving it permission to write back to that board, and allowing it to use a personal laptop are different decisions. The announcement describes app controls, action review and optional device access, not one all-purpose “agent access” switch. For the event launch, start by asking which connected app actually contains the invoice and which output the team wants. A summary in ChatGPT and an email sent to the customer have different consequences even when they begin with the same source document. Initial setup requires the ChatGPT desktop app or a desktop browser. After setup, the user can message the dot in the mobile app; OpenAI also describes Slack and Teams messaging, with texting **coming soon**. Channels make the assistant reachable, not newly authorized to act. At this writing, the primary-dot rollout is underway for **Pro and Business Premium users in eligible markets**; Enterprise users, including Edu and Healthcare, may try the beta when their administrator enables it. A plan name alone doesn't establish that a particular account has access. Specialist dots with company-managed identities are in **focused enterprise pilots**, not an add-on any primary-dot subscriber can simply turn on ([Dots availability and setup](https://openai.com/index/introducing-dots/)). OpenAI's [DevDay recap](https://openai.com/index/devday-2026-recap/) places Dots alongside Codex in the cloud and team collaboration features. They were announced together, but they aren't one entitlement. In particular, cloud Codex availability on other plans does not expand the Dots rollout. The detailed Dots page, not the conference headline, is the place to check who can create one. ## What happens while you're not asking it to act? OpenAI calls the background mode **proactive research**. When you aren't actively working with the dot, it looks for useful next steps using apps you have already connected. Its tools in that mode are restricted to **read-only**. OpenAI explicitly says it cannot send messages, change app content, or control your browser or computer through proactive research ([built-in safeguards](https://openai.com/index/introducing-dots/)). For the imagined launch, this means the agent could look for an overlooked detail in connected material without that research itself sending an invoice. It does **not** mean the dot is a read-only product. Ask it to carry out a task, and a different set of rules comes into play. The official product examples include preparing pull requests and revising project materials; the early tester's invoice was eventually sent with approval. Treat “read-only” as a limit on a particular background mode, not as a blanket description of every way a dot works. Nor should the read-only statement be stretched into a claim that the system is incapable of error or unsafe inference. OpenAI says its safeguards monitor for harmful behavior and may pause or stop work, and separately warns that dots can make mistakes. These are the developer's published controls and cautions, not findings from our own penetration test. A team deciding to connect its customer data should still decide whether the dot needs that data and what subsequent action it is being asked to prepare. ## The invoice crosses a different boundary when someone says “send” Requested actions begin with built-in rules about when a dot can act independently and when it should ask. **Custom Rules** may allow a specific action, require approval, or block it, while built-in safety requirements still apply. An **auto-review** checks actions that could affect accounts or share information against instructions, the Custom Rules and safety requirements. It determines which actions may proceed, which need approval and which must be left to the user ([OpenAI's access and action review](https://openai.com/index/introducing-dots/)). Auto-review is not a human clicking “approve” on every step. Conversely, the ability to write a Custom Rule does not mean the user can override built-in safety requirements. The announcement does not provide a universal action-by-action table for every connected app. It gives one clear human-only example: **changing a password stays with the user**. You cannot infer from that example that all edits require approval, or from the invoice example that every invoice will follow an identical workflow. In practice, a useful instruction for the hypothetical team would distinguish *prepare a draft invoice for review* from *send the invoice*. The first produces something the team can examine; the second makes an external commitment. Inspect the applicable rule and review result for the particular action rather than assuming an approval prompt will appear at a predictable step. If that commitment matters, arrange your workflow so a named person checks its recipient, amount and contents. This is an editorial workflow recommendation, not a claim that we verified a specific Dots configuration. ![Read-only proactive research, reviewed requested actions, and user-only sensitive tasks occupy separate permission boundaries.](/blog/images/dots-permissions-usage-permission-lanes.png "Editorial diagram based on [OpenAI, Introducing dots](https://openai.com/index/introducing-dots/); it does not prescribe approval outcomes for individual actions.") OpenAI says **Activity View** lets users follow background progress and redirect work. That is valuable while an assistant works over several hours, but visibility after an action is not the same as approval before it. Review a consequential output even if the dot has shown its progress: knowing how a draft was assembled does not establish that the result is accurate or that its recipient is correct. The concrete question isn't whether the agent is “autonomous”; it's who may read, prepare, approve and send at each handoff. ## A free conversation can still lead to metered work The invoice isn't the only handoff. Suppose the conversation identifies a bug in the registration page. You ask the dot to start a Codex job to investigate, or manage a ChatGPT Work task to reconcile a list of registrants. If you budget those downstream tasks as if they were just more dot chat, you'll misread the entitlement. OpenAI says the **first dot is included at no extra cost** for eligible Pro and Business Premium users. The plan also contains an **allowance for deeper work**, with extended limits during the first month after launch. The announcement gives no universal numerical quantity for that deeper-work allowance. Additional dots and options to scale a dot's speed or monthly work volume are described as **future** capabilities, not current free entitlements ([Dots get-started section](https://openai.com/index/introducing-dots/)). The accounting language is more precise than “free agent.” **Conversations with your dot do not count toward ChatGPT usage limits.** When the dot starts or manages work in **Codex or ChatGPT Work**, those tasks **count toward usage limits as usual**. A useful distinction is: talk about the bug, let the dot do deeper work, then run a Codex task—three different descriptions, not one unlimited bucket. OpenAI doesn't publish a conversion here from each action into dollars or credits, and we have not run a metered task in an account. ![Dot conversation, deeper work and downstream Codex or Work jobs are different allowance categories.](/blog/images/dots-permissions-usage-allowance.png "Editorial allowance guide based on [OpenAI, Introducing dots](https://openai.com/index/introducing-dots/); no numerical quota or extra-dot price is asserted.") An API token price sheet won't resolve this question. Dots are described through a ChatGPT-plan entitlement, with specific exceptions for downstream products. Whether a plan is enough depends on the work you actually delegate, the limits shown to your account and the terms in force when you subscribe. The first month's extended allowance is not evidence of a permanent limit, and an included first dot is not a promise that an arbitrary number of background jobs will finish without reaching one. ## Decide on the handoffs before you connect the apps There is a practical way to try Dots without pretending to know the behavior of every action. First check your plan, market and, for an Enterprise beta, administrator enablement. Set up a primary dot on a supported desktop surface. Connect only the apps needed for the initial workflow, and leave laptop access as a separate decision. Describe the output you want it to *prepare* before you decide whether it should *send* anything. Then inspect the built-in and Custom Rules around the action that matters. For an external invoice, decide who reviews its recipient and amount; for a coding job, decide who reviews the change and which Codex allowance the job will use. Check the deeper-work allowance in the account rather than inventing a launch-wide quota. These steps won't eliminate mistakes, but they turn a broad claim of being “always on” into a set of choices the team can actually make. Dots' distinctive offer is continuity: an assistant can keep track of a goal and return with useful work while a person is away. The price of that continuity is paying attention to **where research becomes action** and **where conversation becomes another product's task**. OpenAI has published the boundaries above as of October 2. The specific approval outcome and remaining allowance for *your* workflow still have to be checked in the product, not inferred from an announcement. --- ## Sources ### openai.com - [OpenAI's Dots announcement](https://openai.com/index/introducing-dots/) - [OpenAI DevDay 2026 Recap](https://openai.com/index/devday-2026-recap/)