Leave the company AI assistant off until an admin turns it on
On 29 September 2026 OpenAI introduced dots: agents that stay on, each with its own cloud computer, able to keep working when you are not in the chat. The product page says rollout starts on Pro and Business Premium. Enterprise, including education and healthcare workspaces, waits until an admin enables the beta. The help center is stricter: Use dots (Beta) is off by default for Enterprise. Other permissions do nothing until that switch is on. Specialist dots, with their own identity, are a limited pilot. Responsibilities, tools, and how a person reviews the work are defined with the company, not assumed.
That Enterprise default, Use dots left off, is worth copying whether or not you use ChatGPT. What you copy is the switch sitting with an admin, not a requirement to turn on the same product. Day one of an assistant should not include the right to change an order. Write a scope a person can read. It may read answers that were already approved, and it may read order status. It may draft a reply. It may not cancel, refund, change a price, or mail the whole customer list. If that page is not written yet, the admin leaves the switch off. When it is written, the person who can open the admin turns it on. Each employee does not connect their own account to company data.
What differs from a chat box this time is that the assistant stays on. A person closes the window, and work may still continue on that cloud computer. “I will look at the chat before I leave” is not a control. The admin needs to know whether it will still send mail tomorrow, and whether it will still click the back office. OpenAI’s safety note says you choose the accounts, important decisions stay with you, and a dot’s context can be reset. That page leaves permission on the human side. Copied onto your own assistant, it means a write such as cancel, refund, or send-mail waits until a person has seen it. Being awake at night is not a reason for it to press send alone.
Company mail and the back office do not ride a personal login. If a personal plan already has an assistant, and an employee connects shared Gmail or the admin with their own subscription, the switch inside the company workspace does not cover that path. The voice note already said confirms stay on screen, and work mail stays off a personal account. This note adds the workspace: if you have a company plan, use that plan’s switch. If you do not, do not hand the back-office account to any tool that keeps clicking after you go offline. Top admin rights still stay with a person. See models that can click a desktop.
After an admin turns the assistant on, they look at two things a week. Skip model scores. First, the tool list: did someone add payments, or the ads account. If they did, remove that extra tool from the list. Second, messages it has already sent: did one promise a date you did not approve, or a price you did not approve. Fix the wrong sentence in the knowledge set, not only by deleting that one chat. If a buyer already saw the wrong price, a person corrects it on the original channel. Do not wait for the assistant to add a follow-up on its own.
When we wire an assistant for a company, we ask the same four questions, and we do not lock the answer to one vendor. Who approved the knowledge. Who sees a write before it lands on an order or goes out as mail. Who gets the handoff when it is unsure. Whether an admin can turn the whole assistant off. That scope is on the AI section of the service page. Products change. The four questions do not. If a vendor’s admin has no off switch, and no “draft only, do not send,” it is too early to connect the company inbox or the orders.
Two pictures after the rule is clear
kakinang (askai.tw), built with a partner, describes the path as clarify, check, confirm, hand off. When a guest wants to leave a lead or a booking, the screen shows the contents and waits until a person has seen them before it sends. Answers come from knowledge someone already approved. Low confidence goes to the person on duty that shift. That is a confirm before a write, not an extra chat box on the site. Why people still leave after go-live is the older note, after the assistant is installed. A newer model is rarely the missing piece.
The other picture is narrower, and the company switch cannot reach it. On Friday an employee writes: “I connected my own ChatGPT to the support inbox. It replies at night.” The company workspace has no switch that turns off that personal subscription. The person handling it disconnects the support inbox from that personal account the same day. Support for the week is answered by a person. Shared mail stays on tools the company has already reviewed. The personal assistant can stay on that employee’s own errands. It does not stay on the support inbox.
Questions
An employee turned on a personal assistant. Do we care?
What you care about is not whether the assistant is on inside their personal account. You care whether that assistant is connected to company mail, the admin, or payments. A personal plan can be on by default. That switch is not the company’s. Keep company accounts and company data off personal assistants. Do not let a personal assistant read the shared inbox.
Does off by default mean never enable it?
It does not mean never. First write what it may read, what it may write, and who reviews. If the scope is only order status and a drafted reply, an admin may turn it on. If the scope includes changing orders, refunds, or mail to the whole customer list, leave it off. Until that scope is written, the switch stays off.
Is kakinang the same idea?
The public behavior is: answer from knowledge someone already approved, show a booking or a lead before it is sent, and hand off to a person when unsure. It does not go live and then edit your orders on its own. The product site is askai.tw. Off by default, in this note, means that same rule that a person sees a write before it lands. It does not mean you must switch to that one product.
Write the switch and the scope first AI support and agents
Sources: OpenAI, 29 Sep 2026; workspace help; kakinang. Switch labels follow the admin you are logged into today.