Why I'm Not Using ChatGPT Dots (And What I Built Instead)

Since OpenAI announced dots at DevDay on September 29, the most common question in my inbox has been some version of "are you getting one?" Fair question. I run my businesses on an AI staff, I teach designers to build theirs, and OpenAI just shipped a product that looks, from the outside, like the same thing with a cuter mascot.

I'm not getting a dot. Not because dots are bad. Parts of the design are genuinely smart, and I'll give OpenAI credit where it's earned below. I'm passing because of three questions that matter more to a business owner than any feature list: who owns the agent, who owns what it learns about my company, and what happens to my operation if the subscription changes underneath me.

This post walks through what dots are, what they do well, the parts of the deal that didn't make the keynote, and the system I run instead on Claude Code, Vercel, and Supabase. If you're a designer running a firm and wondering whether to hand your back office to a dot, read this first.

Key Takeaways:

• Dots are rented, not owned: your dot runs on OpenAI's cloud computer, inside OpenAI's product, under OpenAI's rules. You can name it. You can't take it with you.
• The memory is theirs too: a dot's own memories can't be viewed, edited, or selectively deleted. The only way to clear what it learned about your business is to delete the whole dot.
• The price is a question mark: the first month is free of plan limits, and OpenAI has said the long-term quota per plan comes "after that." You're building on a number that hasn't been published.
• Owning the stack changes the math: with Claude Code agents, a Vercel dashboard, and a Supabase database, every charter, every approval, and every byte of memory is in a repo and a database I control.
• The right call depends on what you're protecting: if your work already lives in ChatGPT and you want a helper for low-stakes tasks, a dot is fine. If the agent is going to touch your customers and your money, build it on ground you own.

*Affiliate links may be used throughout this article. Read our Privacy Policy for more information.

Table of Contents

What ChatGPT Dots Actually Are

A dot is an always-on agent inside ChatGPT. You give it a name and an avatar, connect the apps you want it to use, hand it an ongoing goal, and it keeps working on that goal between conversations. It runs on GPT-6 Astra, OpenAI's current flagship model, and each dot gets its own cloud computer and browser so it can keep going when your laptop is closed.

You reach it in ChatGPT, by voice call, in Slack, or in Microsoft Teams, and the context follows it across all of them. It connects to more than 4,000 apps through OpenAI's plugin catalog. It inherits whatever ChatGPT already remembers about you and builds new memories from the apps it touches. When it's idle, it does what OpenAI calls proactive research: reading your connected apps looking for things it could help with, so some work shows up before you ask.

Sam Altman's framing was that you should delegate to it the way you would to a high-agency engineer or a chief of staff. The demos backed that up: a dot turning Slack customer feedback into a tested pull request, a dot noticing an unsent invoice and preparing it, a dot cutting podcast clips and writing show notes from a transcript.

What's genuinely good about the design

Two things in dots deserve credit, and I'd say that even if I never planned to use one.

First, the permission model. Every dot has four tiers of rules you set yourself: take action without asking, take action only if you pre-approved that exact action, ask before acting, or hand the whole thing to you. Underneath your rules sits a floor you can't lower: password changes, moving money between accounts, permanently deleting data, and installing software always come back to a human. Before a dot sends an email or edits a file, a separate review system checks the plan against your rules. That is the right shape for an AI that touches a real business, and I'll come back to it, because it's the part of dots I'd copy.

Second, the "proactive mode is read-only" rule. When nobody has given the dot a task, it can look but it can't touch. It cannot send a message, change a file, or drive the browser on its own initiative. That one sentence prevents a whole category of 3 a.m. disasters.

So if the design is this thoughtful, why am I passing? Because none of that thoughtfulness changes who holds the keys.

The Ownership Problem

Here is the sentence that settles it for me: a dot is a feature of a subscription. It is not a thing you possess.

Your dot runs on a computer OpenAI provisions, inside an app OpenAI controls, on a plan OpenAI prices. If OpenAI changes the plan, the dot changes. If OpenAI's safety monitoring decides to pause your dot, it pauses, and their documentation says plainly that monitoring can pause or stop a dot and show you a warning. If you cancel, there is no export button for the agent you spent six months training. You can't move it to another provider. You can't run it on your own machine. The name on the avatar is yours. Everything underneath it isn't.

Compare that to the way you'd think about an actual employee's work product. Their SOPs are in your drive. Their notes are in your CRM. Their inbox is on your domain. When someone leaves, the work stays. A dot inverts that. The agent and everything it has become live on the vendor's side of the line.

The data problem is the same problem

This is where it gets uncomfortable for anyone running a design firm. A dot gets smarter by remembering. It starts with your existing ChatGPT memory and adds everything it learns from your connected email, calendar, documents, and chat. That's the point. That's also the problem, because of three facts from OpenAI's own documentation:

1. You cannot view or edit a dot's memories. There is no page where you read what it has concluded about your pricing, your clients, or your cash position.
2. You cannot selectively delete them. The only way to clear what a dot has learned is to delete the dot, along with its conversations and scheduled tasks.
3. Disconnecting an app does not erase what the dot already learned from it. Pull the plug on your inbox and the knowledge stays.

Then there's training. On personal plans, whether your dot's conversations and work feed OpenAI's model improvement depends on the "Improve the model for everyone" setting. OpenAI also notes that limited human review may happen in safety-related cases even when that setting is off. Business plans are excluded from training by default, which is better, but the memory you can't inspect is still sitting on their servers either way. The documentation even warns that a secret pasted into a message or document may still be visible to the model.

I don't think OpenAI is being sinister here. These are the normal tradeoffs of a hosted consumer product. I'm pointing out that "normal for a consumer product" is not the standard I'd accept for the thing that reads my client contracts.

The Other Costs That Didn't Make the Keynote

Ownership is the big one, but it isn't the only one. Here is the rest of the fine print as it stood two days after launch.

The price is unpublished. Your first dot is included with ChatGPT Pro at $100 a month or with Business Premium. For the first month after launch, dot usage doesn't count against your plan. After that, OpenAI says, the terms for each plan will follow. Additional dots and faster or higher-volume dots are coming "later" at prices nobody has seen. At the same event, OpenAI reshuffled its plan tiers into multipliers, and Pro subscribers were vocal that their tier lost value overnight. If you build a business process on a dot today, you're building on a quota that doesn't exist yet.

One dot per user. At launch you get exactly one personal dot, which means one agent juggling everything you give it. Early testers noticed the cost of that quickly. Peter Yang, who ran dots alongside its competitors for several days, said that because one dot handles several tasks at once, it's hard to tell which one it's replying about. A business doesn't run on one generalist. It runs on roles.

Region locks and gates. Pro users in the European Economic Area, Switzerland, and the UK don't get dots. You have to be 18. You can only create a dot on desktop. Enterprise workspaces have it switched off until an admin turns it on.

It's slow out of the gate. Launch-week reviewers reported around two minutes before a dot even starts a task, and one side-by-side had a dot take 20 to 30 minutes on a job that Claude Sonnet 5.5 finished in five or six. First releases improve, but it's worth knowing what you're paying for today.

It's one more thing inside an app that already has too many. Yang's other observation was that cramming Codex, Work, dots, Spaces, and Pages into ChatGPT is overwhelming. I agree. An always-on agent for your business shouldn't be a tab.

The security backdrop. Four days before DevDay, OpenAI disclosed that its internal research agents had posted 53 users' images publicly. The day before DevDay, it postponed its next flagship model over safety evaluations. OpenAI's own dots documentation says prompt-injection protections reduce but do not eliminate the risk of a malicious instruction causing an unwanted action. None of that is unique to OpenAI. Every agent product carries it. It is a reason to want the agent's blast radius inside walls you control.

What I Built Instead

I don't say any of this from the sidelines. My businesses run on an AI staff today, and it does more than a dot does, on infrastructure I own. Here is the architecture in plain terms.

Claude Code is the workforce. Each agent is a Claude Code agent with a written charter: a markdown file that says what the agent does, what it reads before it acts, what it may do on its own, and what it must bring to me. There are five executive personas, Olive for operations, Mark for marketing, Sally for sales, Frank for finance, and Pete for product, and 20 specialist agents under them. The blog researcher, writer, and webmaster that produced this post are three of them. The agents run on two small virtual servers and one PC in my house, on schedules I set, and every charter lives in a Git repo. If I want to know why an agent did something, I read the file. If I want it to behave differently, I edit the file.

Vercel is the command center. The dashboard is a Next.js app deployed on Vercel. It's where I see the org chart, read what each agent is working on, approve or reject drafts, and watch the numbers. Vercel also runs the scheduled jobs that kick off recurring work. It's a standard web app, which means any developer on earth could maintain it, and I can move it to another host in an afternoon.

Supabase is the brain. Every piece of state lives in a Supabase Postgres database I own: agent runs and their logs, every approval I've ever given, the blog pipeline, the sales pipeline, customer records, revenue, and the media buckets. It's plain Postgres. I can query it, export it, back it up, and if Supabase disappeared tomorrow I'd restore it onto any Postgres server and keep going.

The three pieces talk to each other in one loop. An agent wakes up on its schedule, pulls its context from Supabase, does the work, writes the result back, and if the result needs a human, it pings my phone through Telegram and waits. The dashboard reads the same database, so what I see is what the agents see. Nothing goes out to a customer, onto a website, or into an email broadcast without a row in that database saying I approved it.

Five Reasons Owning Your AI Staff Beats Renting One

I could list twenty, but these five are the ones that would have changed my decision about dots even if everything else were equal.

1. The agents are files, and files are yours

A dot's personality and knowledge are a black box on OpenAI's servers. My agents are text files in a repository. I can read them, diff them, roll them back, copy one to start another, and hand a template to a student. When an agent gets something wrong, the fix is a sentence in a charter, not a support ticket.

2. The memory is readable, editable, and deletable

Everything my agents know lives in a database and a vault of notes I can open. I can see what Sally has concluded about a lead. I can correct it. I can delete a single row. The dots documentation says you can do none of those things with a dot's memory. For a business that handles client information, that difference alone decides it.

3. Roles, not one generalist

A dot is one agent doing everything. My staff is organized the way a business is: a marketing lead who manages a blog writer, a social strategist, and an email agent; a sales lead with four agents working the DM pipeline; an operations lead with inbox, travel, and bill-pay agents. Each one has a narrow job and a clear boss. When something breaks, I know exactly which agent to look at. OpenAI says teams of dots are coming. Mine has been running since spring.

4. The permission floor is mine to set, and I can see it

I said I'd copy one thing from dots, and this is it. Their four-tier rule system is the right idea. The difference is where it lives. Theirs is a settings page inside their product with a floor they define. Mine is written into each charter and enforced in my own code. No agent sends an email, publishes a post, issues a refund, or pays a bill without a human approval that's logged in my database. Finance questions route to Frank before anyone drafts a reply. If I want to add a rule, I add it. If I want to audit every approval I gave in September, it's one query.

5. The cost is visible and the vendor is replaceable

I pay Anthropic for tokens, Vercel for hosting, and Supabase for the database, and I can see every one of those bills line by line. Nothing is bundled into a plan whose allowance hasn't been published. More importantly, each piece is swappable. If a better model ships, I change one setting. If a host raises prices, I move the app. With a dot, the model, the host, the memory, the rules, and the bill are one thing, and that thing belongs to OpenAI.

When a Dot Is the Right Call Anyway

I want to be fair, because "never use dots" is not what I'm saying.

If your work already lives in ChatGPT and Slack, you have no interest in running anything yourself, and the jobs you'd hand over are low-stakes, such as summarizing threads, drafting replies you'll read before sending, or keeping a reading list current, a dot is a reasonable way to get a taste of always-on AI. The approval model is good. The app catalog is huge. You'll learn a lot about what delegation to an AI actually feels like.

Where I'd stop is the line between convenience and operations. The moment the agent is going to read client contracts, see your revenue, talk to customers, or move money, the questions in this post stop being philosophical. Who owns it, who owns what it knows, and what happens when the plan changes become the whole decision. For that work, build on ground you own.

Frequently Asked Questions

What are ChatGPT Dots?

Dots are always-on AI agents inside ChatGPT, announced at OpenAI's DevDay on September 29, 2026. Each dot runs on GPT-6 Astra with its own cloud computer, connects to more than 4,000 apps, and keeps working toward a goal between your conversations. You can reach it in ChatGPT, by voice, in Slack, or in Microsoft Teams.

How much do ChatGPT Dots cost?

One dot is included with ChatGPT Pro ($100 per month) and with Business Premium. For the first month after launch, dot usage does not count against plan limits. OpenAI has not published the long-term per-plan allowance or the price of additional dots.

Do you own your ChatGPT dot?

No. A dot runs on OpenAI's infrastructure inside the ChatGPT product and cannot be exported or moved to another provider. Its memories cannot be viewed, edited, or selectively deleted; the only way to clear them is to delete the dot. OpenAI's safety monitoring can also pause or stop a dot.

What do you use instead of dots?

An AI staff built on Claude Code agents, a Next.js dashboard hosted on Vercel, and a Supabase Postgres database. Five executive personas and 20 specialist agents run on schedules from two small servers and a home PC. Every agent charter is a file in a Git repo, every approval is a row in a database I own, and anything customer-facing waits for a human approval.

Is the Claude Code, Vercel, and Supabase setup hard to build?

It's more work than clicking "create a dot," but it's a standard web stack: markdown charters, a Next.js app, and a Postgres database. All three pieces are widely documented and replaceable. The order I recommend is one operations agent first, then the dashboard, then more roles, which is exactly the build path I teach inside FOCUSED Agentic Designers.

Dots are a good demo of where business AI is heading: named agents with real jobs, working while you sleep, asking before they do anything that matters. I've believed in that picture for a long time, which is why I built it. The difference is that my version is mine. The charters, the memory, the approvals, and the bill all sit on ground I control, and no plan change on someone else's roadmap can take them away. If you're going to hand part of your business to an AI, that's the version worth building.