ChatGPT runs agents on a schedule now. On September 29, 2026, OpenAI shipped team tasks, which run “under the team’s service account,” and dots, which it calls “an always-on agent in ChatGPT.” Add the scheduled tasks that came with ChatGPT Work in July and the Gmail, Slack and GitHub triggers added in August, and ChatGPT does work without anyone pressing enter.
That settles most of the tool side of “AI agent vs ChatGPT.” The question left for an owner who uses ChatGPT every day is what the job around the agent needs. I learned that part by running agents that do real work.
I run hundreds of AI agents across my own companies and client companies. Fifth Nine (fifthnine.com), my company, runs nine of them in its own workspace, with standing jobs that include billing, AR and collections, bookkeeping, corporate filings, the website, and drafting this post.
Is ChatGPT an AI agent?
ChatGPT is both a chatbot and an AI agent, depending on which part you use. Chat is what OpenAI calls “fast, conversational assistance.” ChatGPT Work, scheduled tasks, workspace agents, team tasks and dots are AI agents: they use tools in a loop toward a goal, run in the cloud, and can act on a schedule or trigger.
The cleanest definition I know comes from Simon Willison: “An LLM agent runs tools in a loop to achieve a goal.” He wrote it in September 2025, when he said the word “may finally have” a definition widely enough agreed upon to be useful.
OpenAI’s practical guide to building agents is stricter: “Agents are systems that independently accomplish tasks on your behalf.” The same guide says “Applications that integrate LLMs but don’t use them to control workflow execution… are not agents,” and its first example is simple chatbots.
OpenAI splits its own product the same way. The ChatGPT Work help page says: “Chat is for fast, conversational assistance and everyday questions. Work is an agent designed for longer, multi-step work and finished deliverables.”
By every one of those definitions, my AR (accounts receivable) agent is an agent.1 A chat window drafting an email is a chatbot.
ChatGPT Work vs Chat: what can ChatGPT do on its own in 2026?
ChatGPT Work is the agent half of ChatGPT. Since July 9, 2026 it runs longer tasks, works across connected apps, and runs scheduled tasks that repeat or fire on a trigger. On Business plans, workspace agents and team tasks run in the cloud on a schedule, and actions that send or change data can wait for approval.
OpenAI shipped these pieces between April and September 2026. Each one moves a little more work out of the chat window.
- ChatGPT Work is, in OpenAI’s release notes, “an agent for longer, more involved tasks… You can follow its progress, answer questions, change direction, and approve important actions as it works.” On web and mobile it’s on every paid plan except Free and Go, so Plus subscribers have it. The launch post states the control plainly: “You decide what it can access, when it should check in, and when it needs your approval before taking action.”
- Scheduled tasks “run once, repeat on a schedule or trigger, or monitor for changes,” per the same release notes. OpenAI’s scheduled tasks page says event-triggered tasks “respond to supported Gmail, Slack, or GitHub activity,” and that “An action that sends a message or changes external data may require approval. If approval is required, the task pauses until you review it.”2
- Workspace agents launched April 22, 2026 as agents a whole business workspace can share. The launch post says: “They run in the cloud, so they can keep working even when you’re not.” OpenAI’s Business release notes add that builders can “Add skills, files, and custom MCP servers” and “set safeguards on which actions agents can take for each app.”3
- Team tasks arrived September 29, 2026 on Business and Enterprise. They “run in the cloud under the team’s service account, using the task’s saved instructions and the connections configured for the team.”
- Dots arrived the same day. OpenAI’s dot help page describes “an always-on agent in ChatGPT” with “its own computer and the apps you connect.” You set a rule per kind of action, from “Take action without asking” to “Hand off to you.”4
That is a real agent platform. An owner on ChatGPT Business already has most of the machinery I built my fleet around: schedules, triggers, approval gates and a record of every run.
Custom GPT vs agent: what happens to my GPTs?
Custom GPTs are scheduled to retire on December 11, 2026, and OpenAI is moving them to plugins. A custom GPT is saved instructions and files that you start by hand, and scheduled tasks don’t support GPTs. Workspace agents, which OpenAI calls “an evolution of GPTs,” run on a schedule in the cloud.
OpenAI’s custom GPT retirement FAQ is direct: “Custom GPTs are scheduled to retire on Dec 11, 2026.” It adds: “We’re planning to retire custom GPTs and help creators migrate them to plugins.” Once you migrate a GPT, the original “becomes read-only.”5
The scheduled tasks page lists GPTs among the things scheduled tasks don’t support. Workspace agents are the scheduled path, and the workspace agents launch post says so outright: “Workspace agents are an evolution of GPTs.”
If your team lives in a custom GPT, write down what it does before December. The instructions inside it are the first draft of a job description. Whether it becomes a plugin or an agent comes down to one question: does anyone need it to run without being asked?
AI agent vs ChatGPT: what’s the difference for a business?
For a business, the difference between ChatGPT and an AI agent is the job around it. ChatGPT Chat answers when you ask, and ChatGPT’s agents can run on a schedule with approvals. An agent that holds a job also needs written rules, access to every system it touches, a way to hold exceptions, and proof the work got done.
The third column below is a job design, filled in with my AR agent. You can build that column on ChatGPT Business when everything the job touches sits inside the apps ChatGPT supports. I built mine outside ChatGPT.
| ChatGPT Chat | ChatGPT Work, tasks and workspace agents | An agent that holds a job (my AR agent) | |
|---|---|---|---|
| Who starts it | You, by asking | You, a schedule (paid plans: up to once an hour)2, or a supported Gmail, Slack or GitHub event | A schedule: every Monday at 9am |
| What it can reach | What you paste or connect | Connected apps and plugins, a browser on some websites, and on desktop, apps on your computer; a system with none of those needs a connector someone builds, such as a custom MCP server | Bank transactions and invoices: reads them, and writes back only its attribution note |
| What needs approval | You’re in the conversation | Actions that send a message or change external data can pause for approval | Written rules: it drafts collection reminders, and I send them |
| Exceptions | You handle them | Whatever the saved instructions say; a dot can “Hand off to you” | Held, never guessed: unknown payer, payment larger than the open balance, no open invoice, anything that looks like a refund. Each gets a recommendation and a deadline |
| Proof the work got done | You read the answer | OpenAI: a “Completed” schedule “does not confirm” a run did the work | A dated record of every application; a note counts only once it reads back, and one that didn’t land stays on the weekly list |
| Plan | Every plan | Work: paid plans except Free and Go. Workspace agents: Business, Enterprise, Edu. Team tasks: Business, Enterprise | Runs outside ChatGPT on Fifth Nine’s own agent setup |
The gap that’s left is narrow, and I’d rather name it exactly. Triggers and connections cover the apps ChatGPT supports. Work’s browser can handle some signed-in websites, and on desktop, Computer Use can run steps in your own apps as part of a scheduled task. A team task’s unattended run can’t complete a new app sign-in. For a workspace agent or team task running in the cloud, reaching a system with no ChatGPT plugin, like your own line-of-business software, means a connector someone has to build and maintain, such as a custom MCP (Model Context Protocol) server.
OpenAI’s approvals are gates on single actions. Somebody who knows the business still has to write where the line sits. Team runs also “cannot have separate access restrictions,” so everyone on the team sees every run, which matters for finance work.6
The last gap is proof. OpenAI’s team tasks page says it plainly: “A Completed time-based schedule… does not confirm that an individual run completed the work successfully.”
What does an AI agent need to do a real job in my business?
An AI agent needs six things before it can hold a job: a named job, access to every system that job touches, written rules for what it does alone and what waits for the owner, a rule to hold exceptions instead of guessing, a check that the work was done, and an owner who answers its questions.
I can name all six because every one of them shows up in my AR agent’s record.
A named job. The AR agent’s job is the Monday sweep. It detects incoming client payments, applies each one to the oldest open invoice under a standing policy I approved, records the attribution, and produces an aging digest. That sentence is specific enough to check every Monday.
Its own access, scoped to the job. The AR agent pulls bank transactions and invoices. It can see money arrive. The only thing it writes back to the bank is its attribution note, and moving money isn’t part of its job. For finance work, you decide the access, and read-only is a good default. OpenAI made the same call when it launched ChatGPT agent in July 2025: “ChatGPT is trained to actively refuse high-risk tasks such as bank transfers.”
Written business rules. At Fifth Nine, agents proceed on reversible work by default and report after. Only five kinds of decision need me: any spend or commitment over $1,000, contracts, anything that goes out under my name, legal or government deadlines, and security incidents. OpenAI gives you a gate on each action. Somebody who knows the business has to decide where the line sits and write it down.
When something doesn’t fit, it stops and asks. The AR agent holds four kinds of exception for my decision: an unknown payer, a payment larger than the open balance, a payment with no open invoice, and anything that looks like a refund. It doesn’t guess at any of them. For each one it writes a recommendation and a deadline, and the Chief of Staff agent brings it to me if it’s my call. If nobody answers by the deadline, the recommendation runs.
A check that the work got done. The AR agent records every application in its dated weekly digest. An attribution note on the bank transaction counts only after the agent reads it back. When the bank won’t take the note, the agent doesn’t force it: it says so and carries that line in every digest until it lands. OpenAI’s own team tasks page warns that a Completed schedule doesn’t prove the run did its work. The read-back, and the line that stays on the list when it fails, is my proof.
An owner. That’s me. Agents’ questions go to a Chief of Staff agent first, which answers what it can and brings me the rest.
I wrote all six parts myself, and that writing was most of the work.
What does my AR agent actually do every Monday?
Every Monday at 9am, my AR agent pulls bank transactions and invoices, applies each client payment to the oldest open invoice, records the attribution, and produces an aging digest. Through October 12, 2026 it had run 12 times since July 27: 11 completed and one failed. It costs me less than $200 a month to run.
On August 24, the AI subscription account behind it had hit its monthly spend limit, so the agent never started the work.
The run history recorded why it failed, and that record is the reason to keep one. The one failure was plumbing, which is where I’d look first on any scheduled agent. OpenAI documents the same class of problem for its own tasks: “An unattended run can’t complete a new app sign-in for you,” and “Inactive tasks may pause automatically.”
One thing never happens without me: the agent drafts tiered collection reminders, and I send them. Exceptions wait until a stated deadline, then the written recommendation runs.
This post came from another agent on the same rules. Blog Creator runs three times a week on a schedule, scans news and vendor feeds (37 of them on this run), picks a topic, drafts the post and opens a draft pull request on the website. It can’t publish on its own. A post goes live only after I reply “publish” on that post’s issue, because posting in my name is on the list of things that need me.
When is ChatGPT enough?
ChatGPT is enough when a person starts the work, reads the result and owns the decision: writing, research, analysis, a one-off report. Most small-business AI use already looks like that. In the Federal Reserve’s Small Business Credit Survey, 46% of employer firms reported their business or employees were using AI for work.7
Of those, 83% used it for writing or marketing; 21% used it for process automation. That split looks right to me. Chat work is most of what most owners need, and staying there is fine.
A few signals tell me a job can stay in ChatGPT, either in Chat or as one scheduled task.
- Someone starts it, every time.
- A person reads every output before it goes anywhere.
- It touches only apps ChatGPT supports.
- A wrong answer only costs a redo.
- Nobody else needs to see the run history.
The clearest example is a daily news digest. It needs no company context and no access to company tools, so a ChatGPT scheduled task can do the whole job.
Do I need an AI agent or just ChatGPT? A five-question test
You need an AI agent for a job when it has to run without anyone asking, touches systems outside ChatGPT’s supported apps, needs rules about what waits for you, has exceptions that can’t be guessed, and needs proof it got done. If only the first is true, a ChatGPT scheduled task is usually enough.
Take one job and answer these in order. Each answer maps to one of the six parts above.
- Does it have to happen without anyone asking? If no, use ChatGPT Chat. If yes, keep going.
- Does it touch only Gmail, Slack, GitHub or other apps ChatGPT connects to? If yes, a ChatGPT scheduled task or workspace agent may do it. If no, check for a plugin, the browser or desktop Computer Use first; otherwise someone has to build access to the rest.
- Can you write down what it does alone and what waits for you? If you can’t, you aren’t ready for an agent of any kind. Write that first.
- What happens when the input doesn’t fit? If the honest answer is “it guesses,” keep a person on it.
- How will you know it actually did the work? If the only signal is “Completed,” add a check.
| If the job… | Use |
|---|---|
| Starts when you ask and you read every answer | ChatGPT Chat |
| Runs on a schedule inside apps ChatGPT supports, with simple approvals | A ChatGPT scheduled task, workspace agent or team task (Business plan for the team versions) |
| Touches systems ChatGPT doesn’t connect to, needs written rules, exception handling and a check | An agent built around the job |
Write the job down first. The tool choice gets easy after that.
If you’ve written that job down and want help giving it to an agent, that’s what Fifth Nine’s AI automation work is for.
Sources
How this was written: drafted with AI from my notes and Fifth Nine’s operating records, then edited and approved by me.
Footnotes
-
Anthropic draws a finer line in Building effective agents (December 19, 2024): “Workflows are systems where LLMs and tools are orchestrated through predefined code paths,” while agents “dynamically direct their own processes and tool usage.” It calls both “agentic systems.” ↩
-
Scheduled tasks exist on every plan. OpenAI’s active task limits are “3 for Free and Go, 5 for Plus, 10 for Business and Edu, and 15 for Pro and Enterprise.” Free tasks recur no more than once a day; eligible paid plans go up to once per hour, and event-triggered tasks can run “up to 30 times per hour and 720 times per day.” Source: Scheduled tasks in ChatGPT. ↩ ↩2
-
Workspace agents launched in research preview for Business, Enterprise, Edu and Teachers. OpenAI’s Business release notes later announced general availability in Business, Enterprise and Edu, and say that as of July 6, 2026 workspace agent runs “now use token-based pricing.” ↩
-
Dots are rolling out to Pro users outside the European Economic Area, Switzerland and the UK, and to Business Premium users. Enterprise access is a beta that is off by default. The same page warns: “Your dot can make mistakes, including when following your rules.” Anthropic launched the same idea on Claude, Claude Duet, on October 21, 2026; I compared the two in Claude Duet for small business. ↩
-
Enterprise workspaces with a qualified, approved deferral retire custom GPTs on February 11, 2027 instead, per the same FAQ. ↩
-
From OpenAI’s ChatGPT Learn page on teams: “individual tasks and runs cannot have separate access restrictions,” and “New members can see all earlier runs and generated files.” ↩
-
The 2025 Small Business Credit Survey was fielded September 3 to November 14, 2025 and published in March 2026 as the 2026 Report on Employer Firms. The AI question had 5,248 employer-firm responses; the question about uses had 2,225 responses from firms using AI. ↩
Put AI to work with your team.
A free session using one real workflow from your business. Start with a 30-minute planning call.