ChatGPT Dots Now Follow Custom Instructions to Ask Before Sending Emails
OpenAI updated ChatGPT Dots with custom instructions so your background AI agents actually ask for permission before firing off emails.
Tools · Source: Gadgets 360
What happened
OpenAI just updated ChatGPT Dots to follow custom instructions. You can now tell your background AI agents exactly how to behave. The biggest addition is the ability to force a Dot to show you an email draft before it hits send. You can also tell it to reply in lowercase or stop using corporate jargon. OpenAI says the update makes Dots faster and better at following through on tasks.
You can add these rules directly in the chat while a Dot is working. You can also set them up permanently through the settings menu. On mobile apps, the option lives under Customise and then Behaviour. Web and desktop users will find it directly in the Dot menu. OpenAI claims these agents now respond faster, though they shared no actual performance metrics to prove it.
This update builds on the existing Custom Rules feature. That feature lets you require explicit approval before a Dot changes a shared file or sends a message. Dots are still restricted to eligible ChatGPT Pro subscribers in the US. They run on their own dedicated cloud computers and browsers. They also recently gained support for Apple iMessage and Android RCS so they can text you updates.
Key facts
- September 2025 — When OpenAI originally launched Dots as background AI agents
- US — The only region where Dots are currently being tested with eligible ChatGPT Pro subscribers
Why it matters
Autonomous agents are useless if you cannot trust them. Giving an AI access to your email and shared files is a massive liability for any professional. By letting users mandate approval checkpoints, OpenAI is actively solving the trust gap. Builders need to realize that keeping a human in the loop is not a temporary crutch. It is a mandatory feature for any AI product that takes action on behalf of a user. If your agent cannot pause and ask for permission, it is dangerous.
The second-order effect is a massive shift in how we design agent interfaces. Background agents will constantly pause to ask for permission. They will do this via iMessage or RCS. The notification layer becomes the new battleground for user attention. Apps that seamlessly integrate approval workflows into the daily communication channels of a user will win. Standalone dashboards that force users to log in just to approve a draft will lose.
For builders
Build approval checkpoints into your agents
Do not let your AI send messages or alter files without explicit user sign-off. Users will churn immediately if your tool makes a destructive mistake on their behalf. Make the approval process frictionless but mandatory.
Leverage messaging apps for notifications
OpenAI is pushing Dot updates through iMessage and RCS. You should route your agent approval requests to the places users already spend their time. Forcing them to open a separate app just to click approve is terrible design.
Let users dictate agent personality
Custom instructions let users force Dots to speak in lowercase or drop the jargon. Give your users the same control over how your AI communicates. A rigid corporate tone is a quick way to alienate early adopters.
My take
I never give an AI the keys to my inbox without a leash. OpenAI adding draft approvals proves that fully autonomous agents are still a massive liability. If you build agents, you must build the pause button, or watch your users burn their own bridges.
Original reporting: Gadgets 360. This is my rewrite and opinion.