Claude Cowork and ChatGPT Work: How Far You Get With the Finished AI Tools

Software Selection · 9 min

Illustration: two identical gear motors, one inside a closed housing with a control panel, one built into an open environment of connected building blocks

In a team meeting, an employee shows how he has ChatGPT prepare a quote. The notes from the customer call go in, a usable draft comes out: ten minutes instead of an hour. A few days later a brochure lands on the table promising your company "its own AI agents", at prices that sound more like machinery than like a subscription. Both seem reasonable, and that is exactly what makes the decision hard.

Both run on the same engine.

The same engine in two packages

Claude Cowork by Anthropic and ChatGPT Work by OpenAI are working environments for the desk. You describe a task, the tool works through it step by step, reads documents, produces files and asks when something is missing.

They run on the same language models and the same building blocks that developers use for their own AI agents. When Cowork produces an Excel file, a small program writes the spreadsheet in the background, exactly as a custom-built agent would.

We work with both tiers every day: with the finished tools for office tasks, and with our own agents in our development and operations environment. Seen from the inside, the difference lies in what has been set up around the model.

What the finished tier delivers

For most everyday desk work, the finished tier is the right one. Summarizing meeting notes, drafting a report, pulling an analysis from a spreadsheet, preparing a presentation: such tasks need no setup, no project and no service provider.

Both tools connect to your mailbox, calendar and file stores (through the same interface technology that custom agents use; so the mere connection is not one of the things you need us for). If you want your appointments, mails or notes searched, the subscription covers it.

Prices and feature sets of these products change too quickly for this article to pin them down; check the current state with the vendor. For the decision at hand, the plan you are on plays the smaller role anyway.

Two setups to test before rollout

Keep the expectation for an individual seat modest. In a field experiment by Dillon and colleagues with 7,137 knowledge workers across 66 companies, the people who actually used the provided AI tool spent about two fewer hours per week on email. Beyond this individual time saving, the composition of their work did not measurably change. The Microsoft-affiliated authors studied Microsoft 365 Copilot from September 2023 to October 2024, not Claude Cowork or ChatGPT Work in 2026. The result therefore sets a boundary for an individual seat: a subscription can speed up personal work; it does not organize a shared process.

A Claude Cowork team rollout starts in organization settings, not with a prompt. Decide and record four controls there:

Is Cowork enabled for the whole organization or, on Enterprise, only for selected groups?
May sessions run in Anthropic's cloud, or must they stay local to the user's computer?
May users select Automatically approve mode?
May users permanently allow a connector's write tools instead of approving them for each task?

Leave both approval shortcuts off for the first trial. Then put no more than 3,000 characters of rules that should apply everywhere into organization instructions, and place task-specific guidance and documents in a shared Project. That is the cheapest route to shared context. It remains prompt-level guidance, not technical enforcement. Make a deliberate decision about local sessions as well: their history cannot be centrally managed or exported; Enterprise admins can currently retrieve it only through a beta interface, and cannot centrally delete it.

For ChatGPT, set up a shared working context in this order:

Create one Project for one recurring workflow and add only the approved files and short project instructions.
Share it with the team only after that. Sharing automatically switches on project-only memory, so personal memories and conversations outside the Project stay out.
Before rollout, check whether the workflow needs ChatGPT Work. The current Projects documentation says Work is unavailable with project-only memory. Use the shared Project for team context and Work outside that Project.

Use the same working posture with both tools. Ethan Mollick describes it as a management task: specify the result, material, boundaries and acceptance test, then inspect the output. His practical rule is to allow no more than two correction rounds. If the tool misses the same point again, take over or redesign the workflow instead of supervising a third round. That makes the subscription a bounded first pass, not a colleague nobody manages.

Our own company as the test case

We test how far the finished tier carries on ourselves, because we use both tiers side by side. Research, drafts and analyses run through the same tools anyone can subscribe to. Our development, sales and administration work, by contrast, runs through agents we have set up for ourselves.

The difference becomes tangible in three places. Every one of our agent sessions first reads the same knowledge store: who our clients are, which agreements apply, which rules have proven themselves. On top of that, more than 50 documented working routines record how we review and phrase things; a new session applies them without anyone explaining them. And before a change leaves our company, it passes a fixed review step carried out by a second, independent system. None of this is intelligence. It is organization: the same kind a company gives a new colleague.

A rule from our own work: knowledge that lives in only one tool is lost knowledge. In August 2026 we switched off the built-in memory of our own Claude environment, and beforehand spent one morning reviewing its 172 stored notes from 20 projects. Around two thirds were outdated or long since documented better elsewhere; the rest was moved to where every tool and every employee can find it. Our agents did the reviewing; we made the decisions. Since then, a lasting insight is written down at its permanent place the moment it appears, and not in the history of whichever tool happened to be open.

What the finished tier does not bring

From this experience, the limits of the finished tier can be named precisely. There are four, and none of them disappears with a better model.

First, shared memory. The vendors are moving here: since August 2026 Claude carries one memory across chat and Cowork, and a shared Project gives a team a common body of files and instructions on both platforms. That body of context remains tied to the individual Project, and the product does not decide who removes outdated material or resolves conflicting rules. Which reliable information a task needs, and who maintains it, remains open. That is the question of context management, and it comes before any choice of tool.

Second, a binding way of working. How a quote is structured in your company and what gets checked before anything is sent can now be stored centrally and shared across the company in the vendors' team plans. A stored instruction, however, is a default, not enforcement: Anthropic explicitly calls its organization instruction prompt-level guidance and warns that its priority may vary in rare conflict cases. In an environment of your own, every session applies these routines on its own, changes to them pass a review, and the rules hold no matter which tool is doing the work.

Third, the review step. The finished tools are built for dialogue: a person sees the result and decides. As soon as results are produced regularly or flow directly into systems, you need a safeguard that does not depend on how alert someone happens to be. That means defined checks and, for everything that leaves the company, a human approval at a defined point.

Fourth, operation without a desk. The subscription tools can now handle scheduled and recurring runs themselves, as long as the task works on the connected cloud services; in our own use, however, those runs could not reach the data that lives only on our systems. Sharing such a run also means handing over a copy: each colleague receives their own variant with their own permissions, and the variants drift apart. A jointly operated run with one owner and one log does not exist there. A task that must run reliably every night or after every incoming order needs exactly that: clear permissions (what may be read, what may be written, in whose name) and a log that can be checked later.

How to recognize the limit in everyday work

Whether your company already needs these four things is not decided in a brochure, but in sentences that come up in meetings. These are the signals we listen for when we are asked:

The same correction is being explained for the third time because the tool does not keep it.
Two employees get different answers to the same question because each maintains their own instructions and files.
Results are copied by hand into the ERP, the CRM or spreadsheets.
Nobody can trace which documents the AI saw before a number went to a customer.
A task is supposed to run regularly, even when nobody is sitting next to it.
There are rules that should always apply, and nobody checks whether they were followed.

A single signal is not yet a project. If you recognize three of them, however, every month on the finished tier costs you time, because the tools cannot replace the organization that is missing around them.

Where to start

Start with the finished tools. Give three employees one recurring workflow each for four weeks. Record the expected result, today's handling time and the errors that must not happen before they begin. Then track time spent, correction rounds and missing context. The trial will show which task benefits from the subscription and where a shared process is missing.

Agents of your own become worthwhile once several of the signals above are part of everyday work. The path there starts with the question of which task comes first and which information it reliably needs. The technology follows from that, just as it does when choosing between off-the-shelf software, custom development and spreadsheets.

After those four weeks, decide workflow by workflow: does it remain personal assistance in the subscription, does the team only need a maintained shared Project, or does the process require fixed permissions, checks and joint operation? That classification is more useful than comparing the latest models.