OpenAI's always-on agents keep working after you close the window. Here is how to decide what to hand one, and what to keep.
OpenAI’s dots are always-on ChatGPT agents that keep working between conversations, and OpenAI says they can make mistakes. Decide what to hand one with a five-question score covering recurrence, reversibility, reach, checkability and rules, set Custom Rules by role, and run a 30-day trial scored on use, persistence and impact. We haven’t used dots, and every product fact here is OpenAI’s own claim.
Delegation to a chat assistant usually goes wrong in a familiar way: the task was vague, the output looked fine, and nobody decided in advance what good looked like. An always-on agent raises the stakes on that habit, because it keeps working after you close the window. The first mistakes with tools like this tend to be handover mistakes: a goal with no edges, access nobody scoped, and no moment at which a person looks at the result.
OpenAI announced dots on September 29, 2026 as always-on agents in ChatGPT that take on ongoing work and keep making progress between conversations. Each dot has its own cloud computer, connects to your apps and brings results back for you to review, and OpenAI says it learns your preferences over time.[1] That is all OpenAI’s own description. We haven’t used dots, so this article can’t say how one performs, only how to decide what to give one.
The more useful question is whether a task is one you’d be comfortable running while you aren’t watching. Three things settle most of it: whether a wrong result can be undone, whether the task touches data or accounts you’d keep from a stranger, and whether you could check the answer in a few minutes. The rest of this article turns those three questions into things you can copy.
My position is short enough to say out loud: give a dot recurring, reversible, checkable work first. Anything that speaks for you, spends money or can’t be taken back stays with a person until the dot has earned more.
This is what OpenAI’s launch post, release notes and help centre say as of October 6, 2026. Every line is OpenAI describing its own product, and I’ve flagged where its pages differ.
What we haven’t tried is most of what matters for a decision. We can’t tell you how well a dot’s own review of its actions catches a bad one, how it behaves in week three, or what usage costs after the first month. The scorecard near the end is a plan for finding that out.
If you’re still deciding whether a task suits AI at all, our fit test for any task answers that. This test assumes the answer is yes and asks a narrower question: should the work run while you’re not looking? Score each row 0, 1 or 2 and add them up.
| Question | 0 points | 1 point | 2 points |
|---|---|---|---|
| Does it recur? | One-off. An ordinary chat will do. | Now and then, with no fixed rhythm. | Daily or weekly, with the same shape each time. |
| Can it be undone? | Sent, paid, deleted or published, and you can’t pull it back. | Reversible, with effort or an apology. | It only produces a draft or list that you act on. |
| What can it reach? | Customer, employee, financial or confidential data. | Internal material that isn’t sensitive. | Public pages or your own non-sensitive files. |
| Can you check it fast? | You’d have to redo the work to know. | Checking against the source takes real time. | A few minutes against a source you already trust. |
| Can you write the rule? | “Use judgment.” | A rule with several exceptions. | One sentence: never X, ask before Y. |
Scoring: 9 to 10, start here. 6 to 8, pilot with approval on every outward action. 3 to 5, keep it in a normal chat with you driving. 0 to 2, keep it with a person. In our version, a 0 on undo or reach caps the task at pilot, whatever the total. A Future Factors tool, not an OpenAI feature, and not tested on dots.
If a mistake can’t be undone, a person makes the final move.
A marketing lead who connects a new tool on a Monday usually starts with whatever annoys her most, which tends to be customer replies. Replies score a zero on undo, because a message that has gone out stays out. The competitor-watch job nobody does consistently scores far higher, and it shows you the tool’s behaviour at low risk.
Here are six tasks scored. Treat the scores as our judgment calls. Your data and appetite for risk will move them.
| Role and task | Recur | Undo | Reach | Check | Rule | Total | Verdict |
|---|---|---|---|---|---|---|---|
| Marketer: weekly summary of changes on public competitor pages | 2 | 2 | 2 | 2 | 1 | 9 | Start here |
| Ops lead: Monday status draft from the tracker and Slack | 2 | 2 | 1 | 2 | 2 | 9 | Start here |
| Manager: weekly team update drafted from shared docs | 2 | 2 | 1 | 2 | 1 | 8 | Pilot, approvals on |
| HR partner: morning list of new employee questions, no replies sent | 2 | 2 | 0 | 1 | 1 | 6 | Pilot only if your admin has cleared the data |
| Marketer: replying to customers on social | 2 | 0 | 1 | 1 | 1 | 5 | Draft in chat, you send |
| IT admin: resetting passwords and changing access on request | 2 | 0 | 0 | 1 | 1 | 4 | Keep with a person |
Example scores written by Future Factors to show how the test works. They are not test results. OpenAI itself says certain sensitive tasks, such as changing a password, always stay with you.
The same tool lands differently on each desk. A marketer’s risk is mostly outward, something posted in the brand’s voice. An HR partner’s sits in the data, an IT admin’s in the permissions, and an ops lead’s in a tidy status summary that hides one overdue item.
Custom Rules are where the handover score turns into settings. OpenAI’s launch post names three behaviours, allow, require approval and block, and the help centre lists four settings, so I’d read the three as roughly “without asking”, “ask before” and “hand off to you”. That mapping is my reading, since I haven’t seen the screen. OpenAI says Custom Rules are off by default for Enterprise.[3]
| Role | Allow | Require approval | Block |
|---|---|---|---|
| Marketer | Read public competitor and campaign pages, summarize changes, draft copy into a document. | Anything posted, scheduled or sent outside the company. Ad account changes. Connecting a new app. | Spending, bid or budget changes. Replying to customers or press as you. |
| HR partner | Summarize policy documents. Draft internal FAQ answers for review. | Any message to an employee or candidate. Opening files from the HR system. | Compensation, medical, performance and investigation records. Edits to any employee record. |
| Ops lead | Pull status from the tracker and Slack into a Monday draft. | Reassigning work. Messaging vendors. Creating tickets. | Purchase orders, payments and contract changes. |
| Manager | Assemble prep notes from your calendar and shared documents. | Messages to your team in your name. Calendar changes that involve other people. | Anything about a person’s pay, rating or disciplinary matter. |
| IT admin | Read-only checks, such as summarizing the ticket queue. | Any ticket update. Creating accounts. | Password resets, permission changes and security settings. |
A starting point to adapt, written by Future Factors. Not OpenAI guidance and not tested on dots. If your company has a data policy, the policy wins over this table.
The column people tend to leave blank is Block. Allow and approval describe the work you want done, so they fill themselves in. Block is the list of things that should never happen on a Tuesday afternoon, and it’s quickest to write while you’re still thinking about what worries you.
Anything that speaks for you waits for your approval until you’ve read a month of its drafts.
Rules are a control, not a guarantee. OpenAI says a dot can make mistakes even when it follows your rules, so the next table gives every task a human owner.
Once a task passes the score, write down who owns what, using the split from our guide to writing human-in-the-loop guidelines: what it prepares, what you decide, and how a mistake gets caught.
| Task | What the dot does | What the human still owns | How it gets checked |
|---|---|---|---|
| Competitor watch (marketer) | Checks named public pages on a schedule and lists what changed, with links. | Deciding which changes deserve a response, and anything said publicly. | The marketer opens three of the links each Monday and confirms the change is real. |
| Monday status draft (ops lead) | Pulls updates from the tracker and Slack into a one-page draft. | What counts as late, what gets escalated, and the version that goes to leadership. | The ops lead compares the draft with the tracker’s overdue filter before sending. |
| Employee questions (HR partner) | Sorts new questions into a morning list and flags the sensitive ones. | Every reply, and any judgment about a person’s situation. | Once a week the HR partner reads the raw inbox beside the list to see what it missed. |
Illustrative splits written by Future Factors for tasks scored above. A starting point for your own version.
Every dot gets one named owner and one named reviewer, and they can be the same person.
The “how it gets checked” column is the one that decays. A review that runs every Monday for a month and then “when something looks off” has quietly become no review, because a wrong answer can look as tidy as a right one. A random weekly sample in the calendar holds up better.
This is also where shared practice pays off. One enthusiastic person can set up a dot well. A team trained on its own workflows, with a recurring task and a manager who reads the outputs, tends to end up with rules everyone understands, which is the gap between one person’s experiment and a habit the team keeps. You can start internally with the tables above, and expert-led training shortens the path to that shared habit.
For Business and Enterprise owners, the first question is who is already using one. Pro is an individual subscription, and the help centre says a dot can connect to Slack and to a personal email account.[2] Whether a personal-plan dot can reach a work Slack workspace is worth confirming, because if it can, a dot can arrive the way many AI tools do, through one person and no ticket. Our guide to shadow AI covers why that happens and how to respond without punishing people.
Here is what I’d settle before anyone switches it on:
None of this needs a legal department to start. Our guides to AI governance without a legal department and the AI governance gap cover first rules.
A 30-day trial is only useful if you decide what you’re measuring before it starts. Logins and message counts show that somebody tried the dot, which is access, not adoption. We use three lenses in order: use, persistence and impact.
| Lens | When to check | Question | Record | Keep going if | Stop or reset if |
|---|---|---|---|---|---|
| Use | Weekly | Did the dot do the named task this week, and did I read the output? | Runs reviewed out of runs expected, and your corrections. | Most runs reviewed, corrections shrinking. | Reviews skipped, or corrections not shrinking. |
| Persistence | Day 22 to 30 | Am I still using it without a reminder, and is the review still happening? | Weeks with a reviewed output, and whether anyone had to prompt the owner. | Review is a habit and nobody was chased. | It only runs when someone is nagged. |
| Impact | Day 30 | Did a decision, a number or the time spent change, and would we notice if it stopped? | One specific before and after, in the owner’s words. | You can name what changed, and someone would miss it. | Nothing changed, or checking took as long as doing. |
A Future Factors scorecard for one task and one dot. Not an OpenAI tool. Log any action the dot took that your rules should have blocked, including near-misses.
The weekly shape I’d use is plain. Week one, give it a single task scoring 9 or 10, with approval on every outward action. Week two, check what it got wrong and rewrite one rule. Week three, loosen approval only on read-only steps. Week four, run the scorecard and decide whether the dot stays, shrinks or goes.
Week one starts with a written brief. This one is filled in for the competitor watch, so you can see the standard before you write your own.
| Field | Entry |
|---|---|
| The job | Every Monday, check the public competitor pages on my list and tell me what changed since last week, with links. |
| Sources | Only the pages on that list. No logins and no social accounts. |
| You may do without asking | Read the pages, compare them with last week, and write the summary into a new document. |
| Ask me first | Anything that would post, send, share or edit outside that document. |
| Never | Contact anyone, change an account, spend money, or act on a change you found. |
| When unsure | Stop, and tell me what you were unsure about. |
| Output | A short list, newest changes first, each with a link and one line on why it might matter. |
| Reviewer | Me, every Monday, opening three of the links at random. |
A filled-in example written by Future Factors. Adapt it for your own first task. It has not been tested on a dot.
Four ideas tend to blur here. Training builds capability, so people know how to brief and review an agent. Workflow design gives the dot a recurring task to live in, and reinforcement is the review that continues after the novelty fades. Adoption is what you see if those three hold, and the persistence line catches teams that only got the first. If you want a way to find the right recurring task, our workflow diagnostic is a good place to start.
Pick one recurring task that scores 8 or more, set approval on everything that leaves your hands, name the reviewer, and put the day-30 check in your calendar today.
According to OpenAI, dots are rolling out gradually to Pro and Business Premium users aged 18 and over in eligible markets. Pro excludes the European Economic Area, Switzerland and the UK at launch, and Enterprise, Edu and Healthcare get a beta that stays off until an admin enables it.
We haven’t compared them in use, so we won’t claim a feature-by-feature difference. OpenAI describes a dot as an agent that keeps progressing between conversations. Our guides to ChatGPT agent mode, ChatGPT Pulse and Microsoft Scout cover the neighbouring ideas.
Split the task. Let the dot do the read-and-draft part and keep the risky step, such as sending, paying or changing a record, with a person. That is the Undo Rule in practice, and OpenAI already says certain sensitive tasks, such as changing a password, always stay with the user. Your own Block column should add what is specific to your job, such as spending or compensation records.
Not forever, and not at the same depth. Start with approval on everything that leaves your hands, read the first month of drafts, then move low-risk steps to a random weekly sample. Keep full review for consequential work, as OpenAI itself advises.
For Enterprise workspaces, OpenAI says dots access is off by default until an admin enables the beta. Individual Pro accounts are a different case, since the pages we read don’t describe any employer control over a personal subscription. A clear policy and a place to report personal use are the practical controls there.
This article was written by Hina Mian, Co-Founder of Future Factors AI, and sources were checked on October 6, 2026. Every fact about dots comes from OpenAI’s launch post, release notes and help-centre articles and should be read as a vendor claim, not independently confirmed. We haven’t used dots, so the Dot Handover Score, the starter Custom Rules, the ownership table and the 30-day scorecard are Future Factors planning tools, not tested results and not OpenAI guidance. The example scores are our own judgment calls on example tasks, not measurements. I left out third-party coverage, any pricing beyond the first dot being included at no extra cost, and a feature name that appeared in search snippets but not on an OpenAI page I could open. OpenAI’s pages differ slightly on Custom Rules, which the launch post describes as three behaviours and the help centre as four settings, and plans and limits may change after the first month, so confirm details in your own account.
Weekly tips. Real examples. Practical help for busy professionals.
We care about your data, check out our privacy policy.