What it is
Dots are OpenAI’s always-on agents inside ChatGPT, announced by Sam Altman at DevDay on 29 September 2026. A Dot keeps working after you close the chat. Each one gets its own cloud computer (a Linux environment with a Chrome browser, isolated per user), keeps context across tasks, runs scheduled and recurring work, and makes progress on a goal between conversations. It reaches more than 4,000 apps through the same plugin permissions ChatGPT and Codex use, so an app you have already connected is available to it. You talk to it in ChatGPT on web, desktop and mobile, or in Slack and Microsoft Teams, which are set up from desktop. You can only create a Dot on desktop.
The control surface is Custom Rules. Each supported action sits in one of four modes: act without asking, act if pre-approved (you asked for that action in the prompt), ask first, or hand off to you. Some limits cannot be relaxed: password changes and money transfers always come back to you. A separate auto-review system checks planned steps such as sending an email against your instructions and rules, and runs outside the environment the Dot can change. When a Dot researches on its own initiative, its app tools are read-only.
For an ops team the useful shape is a standing chore, not a one-off task: chase unsigned renewals every Monday and draft the nudge, check the pipeline for deals with no next step and post the list to the team’s Slack channel, or follow up with hiring managers who have not left interview feedback.
Pricing reality
You do not buy Dots separately. The first Dot is included in ChatGPT Pro, which now comes in three tiers at $100, $200 and $500 a month, and in the Business Premium seat, which OpenAI’s help center lists at $100 per user per month billed annually or $125 billed monthly. A Business workspace needs at least two paid seats, so the smallest company setup that includes a Dot is one Premium seat plus one Standard seat. At launch, Pro access to Dots excluded the EEA, Switzerland and the UK. Enterprise, Edu and Healthcare get an admin-enabled beta on custom contracts.
Extra Dots and more speed or monthly capacity per Dot are announced as future paid add-ons with no published price. Chatting with a Dot does not count against ChatGPT usage limits, but Codex or ChatGPT Work tasks it starts do, and allowances are raised for the first month only. Budget the realistic case as $100 to $125 per person who owns a Dot, and expect a second price line once OpenAI prices additional Dots.
What an Enterprise admin controls in the beta
Dots are off by default in Enterprise workspaces. An admin turns on “Use dots (Beta)” in workspace capability settings and controls separately whether Dots can use local computer access, Custom Rules, Slack and Teams, and the cloud browser, network, desktop and password manager. Enterprise model settings do not apply to Dots. During the beta there is no data residency or inference residency, and Dots are unavailable in FedRAMP workspaces and workspaces using EKM. OpenAI has not published a Dot-level audit export; independent hardening guides describe the Compliance API surface as able to revoke access, not reconstruct what a Dot did. Specialist Dots, managed in Microsoft Agent 365 with their own identity and credentials, start as OpenAI-led enterprise pilots with no timeline.
Best for
RevOps managers, CS leads and recruiting coordinators on ChatGPT Pro or Business Premium who own a handful of recurring follow-ups and want one agent to carry them across email, CRM and Slack without building a workflow. It suits a personal chore list, not a team-wide process.
Skip it if the job needs data residency, a full audit trail, or the same logic running for 40 reps: that is a governed workflow, not a personal agent.
Versus the alternatives
Microsoft 365 Copilot and Claude Cowork are the established defaults for agentic work inside a company’s existing AI seat; Manus is the fastest-growing entrant in general-purpose cloud agents. Pick Copilot when the work lives in Outlook, Teams and SharePoint and you need Purview and residency today. Pick Cowork when the work is file- and connector-heavy and you want it on the $20 Pro seat rather than a $100 one. Pick Lindy when you want a purpose-built assistant per process with usage-based pricing from $50 a month. Pick n8n when the follow-up must run identically for every rep, with an execution log you own. Pick Dots when your team already lives in ChatGPT, the chore is judgment-heavy rather than rule-based, and one person owns it.
Watch-outs
- Memory you cannot edit. A Dot builds memories from ChatGPT and connected apps, and you cannot view or delete individual ones; disconnecting an app stops new access but keeps what was learned. Guard: give each Dot one narrow job and connect only the apps that job needs, and delete and recreate the Dot when its scope changes.
- Approval fatigue turns into blanket approval. Four modes per action invite setting everything to “act without asking” after a week of prompts. Guard: keep outbound email to customers and CRM stage changes on “ask first” for the first month, and review the Activity View weekly before moving any action to a looser mode.
- The beta is not compliance-ready. No residency, no published audit export, model settings that do not apply. Guard: in Enterprise, enable Dots for a named pilot group only, keep regulated data (candidate PII, contract terms) out of connected apps until OpenAI documents logging, and route those processes through a governed workflow tool instead.
- Prompt injection is reduced, not solved. OpenAI says its defenses do not eliminate the risk, and a Dot reading inbound email reads untrusted text. Guard: do not grant a Dot both inbox read access and send access without “ask first” on sends.