In 5 seconds
OpenAI’s new Dots are always-on AI agents with their own cloud computers, access to thousands of connected apps, and the ability to keep working in the background. For ecommerce operators, the real question is not whether that is useful. It is what you should let an always-on agent do automatically, what should require approval, and what should stay human-only.
OpenAI used DevDay 2026 to push ChatGPT much further beyond the chatbot model.
The biggest operational change is Dots: always-on agents powered by GPT-6 Astra that can work toward goals 24/7, use their own cloud computer and browser, and connect to more than 4,000 apps through OpenAI’s plugin ecosystem.
OpenAI says Dots can continue working when you are not actively chatting with them, message you when they need input, and use connected systems to research, prepare work, and in some cases take actions.
That is a meaningful change for ecommerce operators.
An AI that waits for a prompt is one thing.
An AI that watches your business all day, notices changes, decides what deserves attention, and prepares or executes work is another.
The most useful part of Dots is not autonomy
It is the control model around autonomy.
OpenAI says Dots begin with rules that determine when they may act independently and when they must ask for approval. Users can create Custom Rules that allow an action, require approval, or block it. OpenAI also says background “proactive research” uses read-only tools, while certain sensitive actions remain with the user.
That is exactly the distinction ecommerce teams should make when designing their own agents.
Do not ask “Can the AI do this?” first.
Ask:
- Should it be allowed to do this automatically?
- Should it prepare the action and wait for approval?
- Should the action remain human-only?
This is the same operating principle we cover in Capability Is Not Permission in Ecommerce Automation.
A practical ecommerce authority model
For a small ecommerce business, a good default is to divide actions into three buckets.
Automatic
Low-risk, reversible, read-heavy work can often run without interrupting the owner.
Examples:
- read inventory levels;
- check order status;
- summarize sales data;
- identify listing issues;
- draft customer replies;
- prepare replenishment recommendations;
- flag late shipments;
- compare supplier records;
- update internal notes or non-public working documents.
Approval required
Actions that can change a live business state should usually stop for a person before execution.
Examples:
- changing a live product listing;
- sending a customer-facing message;
- cancelling an order;
- issuing a refund;
- changing inventory values;
- changing a live price;
- submitting a marketplace appeal;
- creating or changing a shipment;
- publishing content;
- making a change that could affect many records at once.
Human-only
Some actions deserve an even stronger boundary.
For many small businesses, that includes:
- banking activity;
- moving money;
- purchases and spending;
- ad-budget or bid changes;
- bookkeeping entries;
- tax decisions;
- credential or password changes;
- deleting important business data;
- changing security controls that protect the agent itself.
This is not because AI can never perform those actions technically. It is because the consequence is large enough that convenience should not be the only design goal.
Always-on does not have to mean always-writing
OpenAI says Dots can perform proactive research in the background using tools restricted to read-only access.
That is a useful design pattern for ecommerce.
An agent can watch inventory all day without permission to place purchase orders.
It can monitor account health without permission to submit changes.
It can inspect customer-service queues without permission to send every response.
It can track advertising performance without permission to increase spend.
It can monitor accounting data without permission to write financial transactions.
That separation gives you much of the value of continuous intelligence without handing continuous execution authority to the model.
Give the agent a narrow identity, not your whole business
OpenAI says future specialist Dots can have their own identities, credentials, and access to the specific systems needed for their responsibilities.
That is a much safer pattern than giving one agent owner-level access everywhere.
An inventory agent should not automatically need access to accounting.
A listing-quality agent should not need authority over advertising budgets.
A customer-service agent should not automatically need bank access.
The narrower the identity and credentials, the smaller the blast radius when something goes wrong.
We discuss this broader principle in AI Agent Safety for Ecommerce: A Practical Operator’s Guide.
Always-on agents make logging more important
When an AI works only after a human prompt, the conversation itself can provide some context about what happened.
When the agent works continuously in the background, that becomes less reliable.
You need durable records.
For consequential workflows, logs should answer:
- what triggered the work;
- what data the agent read;
- what decision it made;
- what action it proposed;
- whether approval was required;
- who approved it;
- what the target system returned;
- and whether the result was verified afterward.
OpenAI’s Activity View and approval model are useful examples, but the general lesson is provider-agnostic: the more autonomous the system becomes, the more important external evidence becomes.
Do not let OpenAI become your only business memory
Dots are designed to learn from feedback and understand how you work over time.
That can be valuable.
It also creates a portability question.
If your operating knowledge—supplier rules, fulfillment exceptions, approval logic, listing standards, customer-service decisions, and workflow history—exists only inside one AI provider, changing tools later becomes much harder.
Keep durable business knowledge in systems you control:
- operating documents;
- decision logs;
- structured configuration files;
- workflow definitions;
- reusable prompts;
- exportable data;
- and source-controlled code where appropriate.
The AI can use that context. It should not be the only place the context exists.
Dots, Meta Muse, and Amazon Seller Assistant are converging on the same idea
This is not happening in isolation.
Meta is pushing Muse into small-business operations across connected tools. Read Meta Muse Wants to Help Run Your Small Business. Before You Connect Shopify, QuickBooks and Stripe, Read This.
Amazon is opening Seller Assistant to outside AI tools such as Claude through an approved seller-side integration. Read Amazon Blocked Meta’s AI Agent. Now It’s Opening Seller Tools to Claude. Here’s the Difference.
The competitive story is interesting.
The operator story is more important:
AI is becoming an interface through which business systems are monitored and operated.
What ecommerce operators should do before using an always-on agent
- List every connected system. Storefront, email, marketplace, fulfillment, advertising, accounting, bank, and internal files are different risk classes.
- Define automatic / approval / human-only actions. Do this before turning on continuous work.
- Start read-only where possible. Monitoring and analysis create value without immediate write risk.
- Use narrow credentials. Give the agent only the systems and permissions needed for its task.
- Set volume and dollar limits. A valid action can still become dangerous when repeated at scale.
- Require idempotency. Prevent duplicate purchases, messages, refunds, labels, or updates.
- Verify the result. Read back the final state from the system that owns the record.
- Keep durable logs outside the agent. You should be able to reconstruct what happened without trusting the model’s memory.
- Maintain a kill switch. There should be a fast way to stop the workflow without negotiating with the agent that is running it.
Use the right Nyansa Hub resource for the job
If you are deciding whether to connect an always-on AI agent to a live business system, start with the 12-Point AI Agent Safety Checklist.
If an agent or automation already exists and you want to inspect its gaps, use the Ecommerce Automation Safety Audit.
If you are operating recurring or always-on automations, use the Safe Ecommerce Automation Toolkit for the deeper operating layer around permissions, approvals, limits, logs, verification, failure handling and recovery.
If you need to implement those controls in software, use the Safe Ecommerce Automation Coding Build Kit.
For the full governance-to-implementation path, use the Safe Ecommerce Automation & Coding Bundle.
Also from OpenAI DevDay: the tools that agents use are becoming more capable too. Read ChatGPT Plugins Just Became Much More Powerful. Should Ecommerce Operators Build Their Own? for when a repeatable ecommerce workflow should become an installable tool instead of another prompt or checklist.
October 1 follow-up: OpenAI’s MCP Events add the trigger layer to this always-on model. A connected event can start the workflow before the operator notices it, but the event should not automatically grant permission for the downstream action. Read MCP Events Mean Your AI Can Start Working Before You Ask. Ecommerce Operators Need to Understand This.
Cost is becoming part of the architecture decision too. OpenAI’s GPT-6.1 Sol lowers the model-cost side of capable agent workflows. Read OpenAI Just Made Powerful Agents Cheaper. How Ecommerce Sellers Should Recalculate Build vs. Buy. for a practical rent-versus-own framework.
The bigger lesson
Dots make the future of ecommerce automation easier to see.
The valuable agent is not necessarily the one you talk to most.
It may be the one quietly watching the business, catching exceptions, preparing work and bringing you the decisions that actually require you.
But that model only works well when authority is explicit.
Always-on intelligence should not automatically become always-on permission.
Sources and context: OpenAI’s September 29, 2026 Dots announcement and Reuters’ September 29 reporting on OpenAI’s DevDay launch. OpenAI says Dots are rolling out to eligible Pro and Business Premium users, with an Enterprise beta available when enabled by workspace administrators. Features, access and safeguards can change as the product rolls out.