Pat Simmons gave OpenAI's new Dots agent ordinary work: shop for a few things, plan a trip, track a training project, monitor his coding agents, and help with business research. The most revealing moment was not a polished app. It was a failed payment alert that appeared without a specific request to watch invoices. His hands-on review also shows a blocked shopping site, rough output, and unfinished follow-through. That mix is a better starting point than the launch demo.
Dots looks most useful as a persistent coordinator and read-only watcher across connected work. In Pat's session it prepared useful travel research, a training dashboard, agent-status updates, and an unsolicited billing alert. It did not show a completed Amazon purchase, a booked trip, or long-term reliability. Give it one bounded responsibility first, then inspect its evidence and approvals.
Watch Pat Simmons's Dots Test
Credit: Pat Simmons / AI for Mortals, video published 1 October 2026. His companion article, I Gave OpenAI's Dots Real Jobs for a Day, is dated 30 September and provides the follow-up details below. Product details were checked against OpenAI's announcement and setup guide on 2 October 2026. Pat's results are observations from his account, not a controlled comparison.
Three Jobs, Plus the Unprompted Alert
| Job | What the video shows | What it does not prove |
|---|---|---|
| Personal shopper | Amazon blocked the dot's cloud browser. Pat then granted local-browser access, discussed product choices in Slack, and reports cart changes. | No completed purchase is shown. Local access is an additional permission, not a way to ignore a site's rules. |
| Travel planner | The dot compared direct flights and lodging, read review complaints, found an event aligned with Pat's interests, and assembled an itinerary. | Pat praised the research, but the video does not establish a confirmed flight or stay. Check live prices, dates, and booking terms yourself. |
| Ongoing work | It assembled a marathon-training dashboard after a data-access workaround, read Pat's Code Queue project, and scheduled weekday status summaries. | Pat called the dashboard rough. Scheduling a summary is not evidence that future runs stayed accurate or that stale issues were safely closed. |
| Inbox watcher | After Gmail was connected, the dot flagged a failed $22 Vercel payment and $15 of $20 in used monthly credit without being asked to monitor billing. | It did not repair the payment or verify the account state independently. Treat an alert as a lead to check, not as a financial decision. |
Two other examples show the boundary between production and proof. The dot produced a buyer-research document and contact spreadsheet for a family business, but the video does not show verified contacts or a completed sale. It also produced a long newsletter-growth PDF that Pat said he had not read in full; its first summary felt generic to him. Output volume is not a result.
What Pat's Written Account Adds
Pat's companion article makes the handoffs and unfinished steps easier to inspect:
- Shopping: the dot refused to work around Amazon's cloud-browser block. Pat separately enabled access to his local Chrome browser, and the work reached a cart, not a purchase.
- Travel: it caught a complaint about a squeaky Murphy bed at a highly rated Montreal stay and adjusted a route because the Mont Royal cross was under construction. These are useful research details; neither a flight nor lodging was booked.
- Training and operations: when direct Garmin access was unavailable, it used Intervals.icu data to make a rough training dashboard. For Code Queue, it set a weekday 9 p.m. Slack summary and identified 137 issues in progress and 122 in review. Pat asked for proposed closures, not automatic closure.
- Business research: the buyer list contained 136 potential contacts, but the account does not establish that they were verified or converted. His first newsletter-growth report was long and generic until he pushed for more specific advice.
What Feels Different From Another Chat
OpenAI describes Dots as an always-on GPT-6 Astra agent with its own cloud computer, connected apps, and persistent work across conversations. Pat's experience supports the coordination angle: he could message the dot in Slack, return to activities later, and call it to discuss project status. During the call, it agreed to propose issue closures for approval and then sent a Slack follow-up. That handoff is more meaningful than merely generating a page.
The interface still had friction. Pat initially had to reopen the desktop app to connect his local computer, found the activity view confusing, and preferred Slack threads to the mobile view. The dot could receive a call, but it told him it could not initiate the nightly practice call he wanted. He also worried that a growing number of conversations might degrade memory; the video does not test that risk over time.
There is an important distinction between watching and acting. OpenAI says background proactive research uses restricted, read-only tools, while active tasks can request actions under built-in rules, custom rules, and approval checks. Local computer access is optional; it can let a dot use a local browser when its cloud browser is blocked, as Pat tested, but it exposes a different surface of your computer. Connected Gmail, project tools, or a browser should each be treated as a deliberate grant of access. OpenAI also says disconnecting an app does not delete information already obtained by the dot; review deletion controls before connecting sensitive accounts.
Dots vs Grok Bot vs Meta Muse
Pat discusses Grok Bot and Meta Muse, but he explicitly says a deeper head-to-head test is still to come. All three are pitched as agents that can work beyond a single chat. In this video, the differentiator he actually exercises is Dots inside his existing ChatGPT, Codex, connected-app, and Slack workflow. That is a workflow-fit observation, not evidence that Dots is faster, cheaper, or more reliable than the others.
If choosing among them, compare the job rather than the branding: Which accounts can it legitimately access? Can it continue while you are away? Where are approvals visible? Can it show sources and state changes? Does the work stay inside your existing software? Run the same low-risk, measurable task on the products available to you before moving anything consequential.
Who Can Use It, and What Does It Cost?
As of 1 October 2026, OpenAI says Dots is gradually rolling out to eligible Pro users outside the EEA, Switzerland, and the UK. Business Premium is listed across supported ChatGPT regions; Enterprise access is an admin-enabled beta. Users must be at least 18. You create a dot on desktop web or in the desktop app, then can message it on supported mobile surfaces. A personal Pro account in Portugal is therefore not part of the initial Pro rollout.
The first dot is included in an eligible Pro or Business Premium plan. OpenAI's Pro plan guide lists a US$100 tier and a US$200 tier, but that is the plan price, not a separate Dots fee. OpenAI says ordinary conversations with your dot do not consume the normal ChatGPT message allowance; deeper work has its own allowance, and Work or Codex jobs it starts count against those products' limits. Launch-month usage terms are temporary, so check the current plan page before making a subscription decision.
A Safer First Test
- Pick one watch-and-report job. For example, monitor an authorized project inbox for failed-payment notices or summarize a small issue queue once a day. Avoid an entire business or personal life as the first scope.
- Grant the minimum access. Start with read-only data where possible. Keep local-computer access off unless that specific job requires it, and review every connected app's permissions.
- Define the alert. Specify the trigger, source to cite, what counts as urgent, and where the dot should report. Ask it to include a link or record ID so a person can verify each finding.
- Keep consequential actions behind approval. Messages, purchases, bookings, account changes, and issue closures need a human check. Never use a workaround to evade a site's access restrictions.
- Measure the week. Count correct alerts, missed alerts, false alarms, review time, and tasks actually completed. A longer PDF or more notifications is not itself success.
Turn the Workflow Into a Business Idea
Pat's demos suggest a narrow service opportunity: many businesses need someone to watch a specific workflow, reconcile what changed, and prepare the next step. The prompt below helps you find a real buyer and test that job with existing tools, even where Dots is not available.
Find the job behind the agent
Turn a recurring admin problem into a small, reviewable offer.
Video Chapters
| Time | Topic |
|---|---|
| 00:41 | Dots vs ChatGPT Work and Codex |
| 03:24 | Grok Bot and Muse |
| 04:49 | Setting up a dot |
| 09:21 | Personal shopping |
| 11:22 | Talking in Slack |
| 13:44 | Travel planning |
| 16:25 | Marathon dashboard |
| 18:03 | Managing the agent queue |
| 19:11 | Voice call |
| 21:41 | Family-business research |
| 23:32 | Newsletter growth |
| 24:47 | Unprompted payment alert |
| 27:52 | Pat's verdict |
Sources and Useful Links
- Creator test: Pat Simmons / AI for Mortals, ChatGPT Dots review and his full written breakdown. The tests, preferences, and unresolved limitations above are attributed to his work.
- Product and controls: OpenAI's Dots announcement and official setup, availability, and permissions guide.
- Plans: OpenAI's Pro tier help page. Pricing and allowances may change.
- Comparison context: Grok Bot's official launch and Meta's Muse overview; these are product descriptions, not independent benchmark results.