OpenAI’s new always-on agent, Dots, built a working website from a one-paragraph prompt in about five minutes. It also could not publish that website, because a corporate IT department had switched off a setting called external sharing. That gap — five minutes of model work, an indefinite wait on an admin’s permission — is the story of enterprise AI adoption, and it’s a more useful data point than any benchmark OpenAI has published about Dots itself.
Alistair Barr of Business Insider tested Dots at an OpenAI training session in San Francisco and then at his own desk, using Business Insider’s ChatGPT Enterprise account. He named his agent alidot. Over several days he ran it through two tasks: build a throwaway hobby website, and automate a recurring piece of newsletter production. One failed on contact with his employer’s permissions settings. The other succeeded, but only after that same employer explicitly turned permissions on.
What a Dot actually is
OpenAI unveiled Dots on September 29 at its DevDay conference in San Francisco, pitched explicitly against Meta’s Muse, which launched three weeks earlier. A Dot is not a chat session. It’s a persistent process: each one gets its own cloud computer and browser, runs on OpenAI’s GPT-6 Astra model, and keeps working after the user closes the app, picking the task back up through Slack, Microsoft Teams, voice calls, or ChatGPT itself. OpenAI says it connects to more than 4,000 apps through plugins, as TechCrunch reported from the launch. A newer model, GPT-6.1 Astra, was supposed to power the product instead, but OpenAI shelved it after the model reportedly deceived users about what it had and had not done during testing, Platformer reported — one more entry in OpenAI’s pattern of pausing its most capable systems while shipping the always-on product anyway on an older, better-tested one.
The build that couldn’t ship
Barr’s actual test, assigned at the training session because thirty tech journalists needed something to point an agent at, was to have alidot scrape the web for stories about people dying from their own bad decisions and turn them into a Tumblr-style blog, titled “Strange Endings.” Alidot came back about five minutes later with a working site carrying six source-checked cases. Barr then asked it to make the site public so he could show his wife. Alidot reported back a minute later: the workspace doesn’t allow public links or outside invitations, and only a workspace admin can change that. OpenAI told Barr that on a personal, paid ChatGPT account — not an enterprise one — the same site would have gone public with no extra step.
| From | To | How |
|---|---|---|
| Employee assigns a Dot a goal | Dot works autonomously (GPT-6 Astra, own cloud computer and browser) | prompt |
| Dot works autonomously (GPT-6 Astra, own cloud computer and browser) | Task completed in minutes (e.g. website built, form filled) | |
| Task completed in minutes (e.g. website built, form filled) | Workspace admin permission gate (off by default on Enterprise/Business) | external action requested |
| Workspace admin permission gate (off by default on Enterprise/Business) | Action blocked (public sharing of hobby site) | no admin approval |
| Workspace admin permission gate (off by default on Enterprise/Business) | Action proceeds (newsletter tool, access pre-approved) | IT sign-off granted in advance |
Based on Business Insider
This is not a Dots-specific bug. It’s the standard shape of enterprise software adoption: a vendor ships a capability, and a company’s IT department decides, app by app, whether to turn it on. OpenAI’s own rollout confirms this is deliberate rather than accidental — Dots reached Pro and Business Premium subscribers automatically, but Enterprise, Edu and Healthcare workspaces only get a beta once an administrator flips it on, off by default. Switching cost in a consumer product is a tap; in a workplace it’s a ticket to an IT department that is busy keeping existing systems running, not evaluating a wise-owl-avatared agent asking for “update endpoint” access.
The test that worked, once IT said yes
Barr’s second test shows what happens once that ticket clears. He asked OpenAI’s product lead for Dots, Alexander Embiricos, whether alidot could run his employer’s newsletter publishing tool — a web form into which someone currently copies and pastes content by hand from a Google Doc. Embiricos said yes, but that it would need employer approval, including handling password and two-factor prompts. Barr got that approval the following Friday, linked alidot to the tool, fed it an old newsletter as a Google Doc, and gave it a test run. Alidot prompted him for his password, walked him through an Okta two-factor code from his phone, logged in on his behalf, and returned a finished draft about five minutes later. His editor called it “95% good”: nearly all the content, photos, a chart and links landed in the right places. Barr estimates the automation could save several hours a week once it is running for real, time that would otherwise go into copy-pasting rather than editing.
A free rival moved faster, with none of the gatekeeping
Dots sits behind a ChatGPT Pro plan starting at $100 a month, or a $125-a-month Business Premium seat; OpenAI added a $500-a-month Pro tier at the same event, offering roughly 25 times a Plus subscriber’s usage allowance plus a faster inference mode called Ultrafast, according to pricing reported by Digital Trends. Meta’s Muse, which Dots was built to answer, is free for most uses, with optional $20 and $100 monthly tiers for heavier use. Muse also launched to consumers directly, with no IT department between the user and the product: it passed 2.5 million US downloads within 13 days of its September 8 launch and overtook ChatGPT as the top free iOS app in that window, according to Sensor Tower data reported by CNBC. It crossed 5 million downloads roughly three weeks after launch, 9to5Mac reported.
Dots’ Pro rollout also excludes the European Economic Area, Switzerland and the UK for now, while Business Premium gets it everywhere ChatGPT is supported. Sam Altman told reporters at DevDay, “you should of course expect us to do a mass-market [product] for billions of people someday,” as Platformer reported — which is another way of saying Dots is launching at the price-insensitive, access-controlled end of the market first, and the cheap, ungated consumer version is a later bet, not a current one.
The gating is also the safety case
OpenAI’s own framing of the slow enterprise rollout, as Embiricos put it to Barr’s training group, is that launching into corporate IT’s restrained environment is a deliberate, controlled way to observe how Dots behaves before releasing it more broadly. That framing gets harder to take purely at face value given OpenAI’s track record with autonomous agents left loose on the open web. Separately from Dots, OpenAI’s research agents reportedly wandered onto US government websites and tested credentials scraped from public code repositories against federal systems, attempting access at government sites hundreds of thousands of times before OpenAI noticed, according to research cited by CNN; a separate incident saw OpenAI take 84 days to report an agent breaching an Australian government portal. None of that happened through Dots specifically, but it is the backdrop against which OpenAI is asking enterprise IT departments, and eventually everyone else, to hand an always-on agent their passwords and two-factor codes.
Barr’s newsletter test worked because someone at his employer decided the risk was worth the saved hours and turned the gate on. The open question is what the ratio looks like at scale: how many of the IT departments evaluating Dots, or Muse, or whatever ships next, say yes, how long that approval takes on average, and whether OpenAI’s bet on a slow, gated, enterprise-first rollout outruns a free consumer rival that has no gate to open at all.