What Is ChatGPT Dots, and When Does It Fit Operations?

In 60 seconds: ChatGPT Dots is OpenAI’s persistent agent surface inside ChatGPT. A dot can keep working between conversations, use its own cloud computer and browser, coordinate tasks, and communicate through ChatGPT, Slack, or Teams. It fits monitoring, context gathering, and preparing reviewable work. It does not turn an instruction into a reliable transactional process: when a CRM or ERP defines the state of a customer, order, or payment, you still need validation, traceability, and human handoff around that system of record.

What a dot is

OpenAI describes Dots as always-on agents: give one a responsibility, and it can retain work notes and return to the job as context changes. Dots is powered by GPT-6 Astra, but Dots is not another name for the model. Astra supplies reasoning capability; Dots supplies the persistent ChatGPT experience, computer, channels, connections, and controls.

In operations, access to Dots is not the same as deploying a transactional endpoint. Calling GPT-6 Astra through the API does not automatically create persistent memory, channels, permissions, or action review either. The two layers can coexist.

What its cloud computer means

Each dot has its own computer and browser in the cloud. It can retain files, software, and sessions and continue working while your device is off. When a site requires authentication, the dot can request a private sign-in or hand browser control to you. Some sites block cloud browsers or ask for additional verification, and sessions may expire.

The cloud computer is separate from your device and personal browser sessions. You can connect a local computer to expose compatible files, apps, or tasks, but it must stay online with the ChatGPT app open. You can also connect plugins. Each plugin keeps its own account and permissions: adding Slack as a messaging channel does not connect an inbox or CRM.

In practice, the cloud computer is useful for research, documents, analysis, and software work that can progress outside a conversation. When a stable API exists, a contract-based integration with structured responses is usually easier to observe than a sequence of clicks. That matters most in a critical workflow.

Proactive does not mean unlimited autonomy

A dot can decide when to pause and resume a responsibility, create background tasks, and contact you when a decision is needed. Work that must happen at a fixed time needs a saved schedule; connecting a source or writing a cadence in a document does not create that schedule by itself.

OpenAI also separates proactive research from actions. A dot can read authorized information in connected apps and keep private notes to identify useful changes. That research cannot send messages, modify the app, or control a computer. Any follow-up action goes through the permissions and reviews that apply to the work.

This gives B2B operations a sensible starting pattern:

1. observe authorized sources;
2. detect a change or exception;
3. gather evidence and propose the next step;
4. request a decision when commercial, financial, or data impact is involved;
5. execute only the bounded action and record the result in the right system.

Permissions, auto-review, and custom rules

Before an action can affect an account or share information, Dots runs an automatic review against the instructions, permissions, custom rules, and safety requirements. The result can let the action proceed, request approval, or hand the step back to a person.

Custom rules offer four behaviors for a defined scope: act without asking, act when explicitly told, ask before acting, or hand the action to you. A rule can require approval before contacting a customer or reserve shared-file deletion for a person.

Keep three limits in the design:

• a rule is an instruction the agent tries to follow and may misinterpret;
• a rule does not grant app access or expand app permissions;
• pausing or deleting the dot does not undo messages or changes already completed.

The Activity view lets you inspect tasks, results, files, errors, and pending requests. It is useful agent supervision. It is not a substitute for a business audit log tied to the customer, opportunity, or order ID and the person who approved an action.

Decision matrix: where Dots fits and where more system is needed

Operational need	Dots fits	Use a CRM/ERP-integrated agent or custom service
Watch sources and report meaningful changes	Yes, when sources are connected and notification criteria can be reviewed	When every event must be processed once with an SLA, retries, and a technical alert
Prepare a brief, proposal, or response	Yes: gather context and leave a draft for review	When pricing, credit, stock, or contract rules require deterministic validation
Coordinate work across ChatGPT, Slack, and Teams	Yes, as a contact point for the same dot	When the channel needs dedicated queueing, routing, hours, consent, identity, and traceability
Update opportunities or orders	Only with bounded fields, least privilege, and approval where needed	When the CRM or ERP is the source of truth and needs schema validation, idempotency, concurrency control, and operational rollback
Send, pay, delete, or grant access	Only within a narrow scope and with explicit human review	When the action is irreversible, regulated, or requires separation of duties
Hand work to a person	It can request a decision and supply context	When the case must enter a queue with an owner, priority, deadline, state, and mandatory evidence

Dots is strong in the semi-structured work around a process: reading, relating, summarizing, preparing, and coordinating. A custom service becomes necessary when the process needs guarantees the system can verify. The dividing line is the system that owns state and the person accountable for each transition.

Three useful pilots for a B2B team in Latin America

1. Daily opportunity brief

Ask the dot to read permitted opportunities, recent notes, and connected conversations. The output is a short list of accounts without a next step, risks, and follow-up drafts. A salesperson confirms every send, and the CRM retains stage, owner, and activity.

2. Customer meeting preparation

The dot gathers the agenda, recent agreements, open tickets, and authorized documents. Require a source for each fact and a flag for contradictions. A person decides which commitment, timeline, or price to communicate.

3. Operational exception radar

Connect an incident source or operations channel and ask the dot to group changes by order, account, or project. It can propose a priority and owner. The ticketing system or ERP keeps official state, and a human queue handles ambiguous cases.

All three pilots produce a reviewable artifact while authority stays in the system of record. They can also be measured with your own data: coverage, corrections, time to decision, escalations by reason, and actions that reached the right system. You do not need invented ROI to learn whether the workflow removes search work without creating new errors.

Errors to expect

A serious rollout needs observable failure cases and a defined response for each one:

• a connection expires or a website blocks the cloud browser;
• the dot uses incomplete context, a stale note, or the wrong conversation;
• it treats a commercial exception as a standard case;
• a task completes technically, but the result does not reach the expected person or system;
• an action has correct content but the wrong customer, channel, or timing;
• pausing the main task leaves delegated or scheduled work active.

Design an explicit exit for each one: reconnection, required data, a needs-review state, an idempotency key, a post-action check, a human queue, and a shutdown procedure. OpenAI notes that a completed run does not itself confirm that the intended result was achieved or delivered, and that a dot can make mistakes. The user remains responsible for reviewing consequential work.

Checklist before giving it a real responsibility

1. Choose a specific outcome and name the system that owns official state.
2. Connect only the required sources and actions; start with reads and drafts.
3. Define which decisions it can make, which need approval, and which always go to a person.
4. Put business IDs in outputs and retain source, action, and approval evidence.
5. Test missing data, expired connections, duplicates, malicious instructions, and unavailable systems.
6. Review Activity, delegated tasks, and schedules separately; Pause does not stop everything.
7. Give the human queue an owner, operating hours, and a response target.

If the pilot needs many manual exceptions, do not add more autonomy. Narrow the scope, improve the data, or move stable logic into a custom integration. The copilot-or-autopilot framework helps assign autonomy per action, while the GPT-6 Astra business guide covers the model that powers Dots.

Availability and verified sources

Availability is gradual. As of September 30, 2026, official documentation lists access for users over 18 on Pro 100, Pro 200, and Pro 500 outside the European Economic Area, United Kingdom, and Switzerland; Business Premium and Enterprise are rolling out worldwide. Dots is off by default in Enterprise and must be enabled by an administrator. Confirm the account, region, and workspace policy before designing a process around the feature.

Official sources checked on September 30, 2026:

• Meet dots: definition, GPT-6 Astra, channels, cloud computer, controls, and availability.
• Tasks and memory: persistent work, schedules, context, and read-only proactive research.
• Computers and apps: sessions, local access, plugins, authentication, and cloud-browser limits.
• Control your dot: Activity, action review, custom rules, stopping work, and control limits.

OpenAI may change availability, plan names, controls, and connections. Confirm those pages and the effective configuration of your workspace before rollout.