Would you hire an AI that never clocks off?

OpenAI held its developer conference on September 29 and the announcement that matters to a business owner is not a model. It is a colleague. OpenAI calls them Dots: agents that live inside ChatGPT, run on the GPT-6 Astra model this column covered three weeks ago, and, in OpenAI’s words, “can work towards your goals 24/7.” Each Dot gets its own computer in the cloud with its own browser. It can be reached from the ChatGPT app, from Slack, and by voice call, and it connects to more than 4,000 apps through plugins, including your email and calendar. It remembers what you told it last week, learns from your feedback, and when you have not given it anything to do it does read-only research on its own initiative through the apps you have connected. The first Dot is included at no extra charge on the Pro plans and on Business Premium seats. It is not on Free, Go or Plus, it is not yet available to Pro subscribers in Europe or the UK, and OpenAI says more Dots will be available later. The rest of the conference matters less to this column but is worth a line. GPT-6.1 Sol arrived as a cheaper model at a fifth of Astra’s price per token and is available to Plus subscribers in the Work and Codex modes, and a new Pro 500 tier at US$500 a month appeared for very heavy users. If you did the two-hour bake-off from Week 15, Sol is the one to add to the table.

An always-on AI agent inside ChatGPT on GPT-6 Astra, with its own computer and browser in
the cloud, 4,000-plus apps, and memory of what you told it. It can work towards your goals 24/7. (Branded graphic by PGH Consulting, LLC)

What a Dot is, in finance terms The cleanest way to think about a Dot is as a junior hire with three unusual properties. It never leaves the office. It keeps the context of everything you have worked on together. And it will do exactly what you permitted, at three in the morning, with nobody watching. That last property is the whole risk. A member of staff who makes a mistake at three in the morning is usually asleep. An agent that has been given standing access to your inbox, your shared drive and your accounting software is not. OpenAI has clearly thought about this, and the controls are the most important part of the launch. When you set a Dot up you choose which apps it can reach. Then you write what OpenAI calls Custom Rules: for any category of action you can allow it, require your approval first, or block it outright. Anything that touches an account or shares information goes through an automatic review before it runs. When a Dot is working on its own, it is limited to reading: it cannot send a message, change content in an app, or drive a browser or computer without you. Password changes always stay with you, and for supported sites a saved password can be used without the model ever seeing it. Your own device is not connected unless you deliberately connect it. And OpenAI’s own caution is short: “Dots can still make mistakes, so always review consequential work.” Read together, those controls describe a decision a business owner has to make before the Dot does anything. Not “should I get one,” but “what would I let it touch.”

Useful read-and-draft tasks for a finance team: summarise the inbox, pull and organise supplier
invoices, check reports for gaps, manage the calendar, monitor updates. (Branded graphic by PGH Consulting, LLC)

The three settings, and what they mean for money Allow, approve, block. Every job you might give a Dot falls into one of those three, and it is worth deciding them in that order. Allow is for reading and drafting. Summarise the inbox at seven each morning. Watch the shared drive for new supplier invoices and pull them into a list. Draft the reply to a customer query and leave it in the drafts folder. Check whether last week’s receipts match the sales report and flag the gaps. None of that moves money, changes a record or reaches a third party, and all of it is useful.

Allow reading and drafting. Approve anything that leaves the building: emails, invoices,
journals, bookings. Block payments, bank logins, payroll changes and cards. (Branded graphic by PGH Consulting, LLC)

Approve is for anything that leaves the building. Sending an email to a customer. Creating an invoice. Posting a journal. Booking anything. The Dot can prepare it, and you press the button. That costs you about ten seconds per action and buys you the thing a control is for: a human looked before it went. Block is for money and credentials. Payments, transfers, bank logins, payroll changes, anything on a card. This column has argued since Week 12 that an agent with access to money and email needs a human approval before it acts, and for a first Dot the safer position is stronger than that: it does not get the access at all. The permissions register from Week 16 has a third line for exactly this, and a Dot is the first product that will ask you to fill it in. A first job that is worth the subscription The temptation with an always-on agent is to give it something dramatic. The better first job is dull, high-volume and read-only, which is the same rule this column gave for choosing what to automate in Week 14. Supplier statements are a good candidate. Connect the mailbox where they arrive and the folder where invoices are kept, and set the rule: every day, list new statements, match each line to an invoice on file, and flag anything that does not match. No sending, no posting, no payment. In a month you will know two things: whether the Dot is accurate on your data, and whether the accounts payable clerk’s afternoons just got shorter. A second candidate is the customer-payment chase. The Dot reads the aged receivables export and the inbox, drafts a reminder for every account past thirty days in your tone, and leaves the drafts for a person to send. The judgment about who to chase and how hard stays human. The typing does not. Score both the way you scored the model bake-off: correct, followed the instruction, how much editing. If the Dot is not earning its place on read-only work, it should not be promoted to approve. What the launch does not settle A few plain limits. It is not cheap. The first Dot is included, but the plans that include it start at the Pro tier, several times the US$20 most small businesses pay, or at a Business Premium seat. For a sole trader the arithmetic may not work yet, and the read-only jobs above can be done today on a cheaper plan with a scheduled task and a person pressing run. It is not in Jamaica in any special way. Dots follow ChatGPT’s supported regions, so a local Pro or Business Premium subscriber should be eligible, but rollout is gradual and OpenAI has said some accounts will wait days. Check before you budget. It is a personal agent on a personal account unless you put it on a business one. A Dot set up on an owner’s Pro subscription learns the owner’s inbox and the owner’s files. The Business Premium version sits inside a workspace an administrator controls, which is where company data belongs, and which is the version to use if staff will be involved. And it is new. OpenAI’s own safety monitoring can pause or stop a Dot when it detects a concern, which is reassuring and also a reminder that the company expects concerns. The three-in-the-morning test applies to the vendor too: a product launched on Tuesday has not yet had a bad Thursday.

A Dot will do exactly what you permitted, at three in the morning, with nobody watching.
Decide the permissions before the first task, not after the first mistake. (Branded graphic by PGH Consulting, LLC)

What to try this week 1. Before you look at the plan page, write the three lists: what a Dot may do on its own, what needs your approval, and what it must never touch. If money is anywhere but the third list, start again. 2. Pick one read-only job with real volume, supplier statements or the receivables chase, and define it in a paragraph a new hire could follow. That paragraph is the Dot’s instruction. 3. If you subscribe, connect the smallest set of apps that job needs and nothing else. Add access later, when it has earned it. 4. Use the Activity view for the first two weeks and keep a tally of what it got right and what it got wrong, the same ten-cell scorecard as the model test. 5. If your staff will use it, set it up on a Business Premium workspace, not on someone’s personal plan, and record the Custom Rules in your AI permissions register. Dots, plan inclusions, regional availability and controls are as described by OpenAI at launch on 29 September 2026 and reported in the coverage cited in the supplementary notes; they are rolling out gradually and may change. Prices quoted are the published United States figures. An agent that works while you sleep does not remove the need to check its work when you wake up. Peta-Gaye Hardy is the founder of PGH Consulting, LLC, where she helps finance and operations teams adopt AI in practical, low-risk ways. She writes the weekly AI in Finance & Business column and is based between Jamaica and the United States. Learn more at www.pghconsultinggroup.com. Follow on Instagram and YouTube @pghconsultinggroup, and connect on LinkedIn at linkedin.com/in/peta-gaye-hardy. Disclosures: This article is informational and does not constitute investment, tax, legal, or accounting advice. Readers should consult a qualified professional before acting. Product features, plan inclusions, prices and regional availability are as published by OpenAI at the time of writing and are subject to change; readers should confirm the current position on the vendor’s own pages before subscribing. AI tools can produce errors, and every figure, clause, or claim they produce should be verified against a source before being shared or acted upon. The author has no commercial relationship with OpenAI or any product mentioned and was not compensated by them. All product names, logos, and trademarks are the property of their respective owners and are referenced for editorial purposes. The examples described are illustrative and do not depict any real business.