Dots, the always-on assistant OpenAI released in September, grew out of its coding agents. OpenAI's own view is that little new technology went into it, and that the work was in "how you frame it to the user and how it feels."
The wrapping does real work
With the Codex app, an agent could keep working in the background, but it ran on your own computer. OpenAI's product lead was pleased when his wife started leaving her laptop on all night so her agents could carry on. A dot has its own cloud computer, so nobody has to leave anything running at home, and you talk to one assistant instead of keeping track of a separate thread for every task.
That matters because most people will not manage a team of agents. They will keep one colleague who knows what they care about. OpenAI's own legal, finance and recruiting teams moved to Codex as their main AI tool around April, so the appetite among people who do not write code was already there. What was missing was something they could picture themselves using.
What it looks like for a small business
Take a shop owner who supplies local restaurants. At 6 in the morning her supplier emails that Thursday's delivery will arrive a day late. Without a dot, she finds out at 9, when a restaurant phones to ask where its order is.
With a dot connected to her email and calendar, the dot reads the supplier's message, sees the standing Thursday order in her calendar, and drafts a note offering the restaurant a Friday delivery. The background watching is read-only, so the dot can only propose the note. It goes out when she taps yes on her phone.
Written by hand, that note is ten minutes of work if she notices in time. Approving it takes about half a minute. The real saving is the three hours between the email arriving and the phone call, in which she can warn her customer instead of apologising to him.
The limits are permissions, not ability
Dots can act on their own only inside rules the user sets. Custom Rules let you allow an action, require approval for it, or block it, and certain tasks, such as changing a password, always stay with you. The controls are still coarse. In TechCrunch's testing, the assistant would not accept read-only access to cloud drives, and OpenAI's product lead says the controls need to get much finer.
The work it does also costs something. Chatting with a dot does not use up your plan's allowance, but the tasks it runs do, so a dot that watches constantly is a dot that spends constantly. At launch it is limited to the plans from $100 a month and up, and Pro users in the European Economic Area, Switzerland and the UK were left out.
Payments are the next permission. OpenAI is in talks with payment processors, which is a conversation and not a launch. If it happens, the shop owner's dot could pay the supplier's late-fee invoice as readily as it drafts the note, and the rule she sets for how much it may spend without asking will matter more than anything else on the settings page.
Before connecting anything, decide what you would let a new assistant do without checking with you, and what you would want to approve first. A dot starts out asking often, and Custom Rules is where you write that down once instead of answering the same question every day. Dots has been available since September, and the rules you set decide whether it is any use to you.