ChatGPT Work in practice: automation and marketing workflows
After Codex, OpenAI has taken the next step: ChatGPT Work is a dedicated work agent inside ChatGPT that doesn't chat with you, it gets tasks done: it creates files on your machine, connects to your systems and runs your routines on a schedule. We put it through its paces and collected what a business should actually take away from it.
What is ChatGPT Work, and how is it different from regular ChatGPT?
The core difference is purpose. The familiar chat is for asking, brainstorming and getting explanations. Work is for getting finished output back: whatever you start there is not a conversation but a task, and at the end of it there's a completed file sitting on your machine. In the desktop app a dropdown switches between Work and the developer-oriented Codex, and for quick thinking there's a quick chat: once a train of thought has taken shape there, one click attaches it to a task and Work carries on with the full context.
A practical detail: the two modes draw on separate usage allowances. Work has its own limit measured in a several-hour and a weekly window, while regular chat has its own. If you keep quick questions in chat and real work in Work, you manage both budgets at once. You pick the model with a slider between the faster-and-cheaper and the smarter end; the middle setting runs the GPT‑5.6 Sol medium model, which we covered in a separate article.
Projects and files: the work lands in your own folders
Work shows its real strength in the desktop app, because that's where it can reach the file system. A project here is not a label but a workspace bound to a real local folder: you point it at an existing folder, and from then on the agent works from its contents and saves results right back into it. In one of our tests it built an Excel summary from the files in a budget folder, then on a follow-up request a PowerPoint visualising the spending, and both appeared instantly in the folder, ready to open and edit.
The safety brake is built in: permission levels control what the agent may touch. The default asks for approval before any external file operation or internet access, the looser level only stops at steps it judges risky, and full access requires a deliberate decision. In a business setting the first two are the sensible choice.
Plugins: wiring it into your real systems
Beyond local files, the second step is connecting external systems. The plugin list includes Gmail, Google Drive, Notion, Slack, QuickBooks, Canva and Figma: the agent can read, search, create and modify in them, depending on each plugin's permissions. With a separate extension it can also drive the Chrome browser: in one of our trials it worked through a prepared list of search queries across Google Shopping results and built an analysis document from competitors' ad patterns.
The Gmail integration is the most striking: the agent doesn't just read the body of an email, it processes attachments too. In one workflow we tried, it took the CSV attachment of a scheduled Google Analytics report email, produced a weekly performance summary and replied to the same thread with the finished report attached. An important and reassuring detail: it prepares draft replies but never sends anything on its own; hitting send remains your decision.
Skills and templates: how the output stays consistent
The skill is one of Work's most valuable concepts: a repeatable workflow packed into a portable format that the agent executes again and again at the same quality. It can live at three levels: bound to your user account, to a project or to the system. If a presentation comes out with a style you like, a single sentence turns it into a skill, and from then on every deck carries the same look. The template-creator plugin does the same with existing documents: in our case a client-onboarding brief became a reusable template that the agent filled in completely even from a messy pile of source material.
Marketing workflows worth trying
Attach a market-research report, shape the storyline in quick chat, then get a strategy deck built to your own brand skill, all in a single task.
A weekly ad sweep via browser control, feeding a continuously updated playbook document.
Images generated for the coming weeks of posts in your Notion calendar, uploaded to Google Drive, with the calendar entries updated automatically.
A site built from the campaign brief and brand guideline stored on Drive, hosted on OpenAI's servers, with on-brand generated imagery.
The Sites feature deserves attention on its own: finished pages can stay private to you or be shared publicly, so you can even build an internal dashboard that refreshes daily and check it from your phone.
What to watch out for
Testing also surfaced a few important limits. Scheduled tasks only run while the computer is on and the desktop app is open. The web version handles cloud plugins without trouble but cannot reach local files; the mobile app reaches your desktop projects over a remote connection. And most importantly: the output is not flawless. There were runs where the agent skipped the defined logo, or left out entries of a content calendar until the date range was spelled out explicitly. The lesson is simple: precise instructions, a narrow mandate, and human review of every output before it reaches a client.
What does this mean for your business?
The barrier to entry keeps dropping: ChatGPT Work is built specifically for non-technical users, and in our experience it sits comfortably alongside the similarly positioned Claude Cowork, each with its own strengths. The more important question than picking a platform is whether the usage grows into a system: without permission rules, shared templates, scheduled routines and review checkpoints, even the strongest tool remains scattered experimentation. Fit it into your processes, though, and a good part of the weekly reporting, competitor watching and content production comes off your team's shoulders.
Curious where a work agent like this would pay off in your team's processes? Let's find out together on a free 30-minute consultation.
Free 30-minute consultationFrequently asked questions
How is Work different from regular ChatGPT?
Chat is for questions and brainstorming; Work runs tasks and hands back finished files. Their usage allowances are separate, so it pays to split your work between them deliberately.
Do I need to know how to code?
No. Work is the interface built for non-technical users; the developer Codex lives in a separate mode of the same app.
Which systems can it connect to?
Via plugins: Gmail, Google Drive, Notion, Slack, QuickBooks, Canva and Figma among others; with an extension it drives Chrome, and in the desktop app it works with your local folders.
Do scheduled tasks run on their own?
Daily and weekly routines can be scheduled, but for tasks touching local files the computer must be on and the desktop app open.
Last updated: