Skip to content
Auf Deutsch ansehen

AI · AI at work

ChatGPT Work for teams. From a brief to a checked deliverable

What ChatGPT Work can do with files and apps, how local and cloud access fit together and how to measure the effort and quality of a real assignment.

By Boaz Lichtenstein Prefer us on Google

Article image: ChatGPT Work for teams. From a brief to a checked deliverable
In this article
  1. Define what a finished report should contain
  2. Which environment can see which data?
  3. Choosing a model and measuring usage
  4. Give the spreadsheet a five-minute acceptance test

As of 4 November 2026. A good first assignment for ChatGPT Work is the weekly report that takes an hour every Monday. It has identifiable inputs, a calculation someone can check and a particular audience. That makes it possible to tell whether the agent saved work or handed the team a convincing draft that still needs rebuilding.

Define what a finished report should contain

Give Work an anonymised order export and last week’s report. Ask for an editable spreadsheet showing revenue, refunds and week-on-week change. Put the cut-off date in the file and flag missing data in a separate row. Ask for a draft message to the team as a separate deliverable. A calculation error can then be caught before it reaches the sales team’s inbox.

Work can use files and approved apps across several steps and deliver a file you can keep editing. A question about a definition rarely needs this machinery. With the weekly report, you have formulas, sources and repeated steps that make a meaningful acceptance test possible. OpenAI’s introduction to Work

Which environment can see which data?

A cloud task does not inherit your laptop’s files or its signed-in browser session. Context may come from an upload, an authorised app or separately approved access to a connected computer. A step that needs a file held only on that computer requires the device to be reachable. Running a step locally does not promise that all processing stays on the device. OpenAI’s execution guidance

For the report, an export with pseudonymous order IDs will do. Connecting an entire mailbox adds exposure with little analytical value. For a shared connection, check what the connected account itself is allowed to read. Try a record outside the intended scope when testing access. Connected-app permissions

Choosing a model and measuring usage

The pricing page also lists Work and Codex for Free and Go. They share usage under the applicable ChatGPT plan. The GPT-5.6 family featured over the summer tells only part of the story today. OpenAI lists GPT-6 Luna for focused tasks, GPT-6.1 Terra for writing and visual work, and GPT-6.1 Sol and Astra for more demanding assignments. The options you see depend on the plan, client, rollout and administrative settings. Plans and allowances · Model guide

For three metrics and a short commentary, start with the available default. Run the same assignment twice with identical inputs. Note the actual allowance used, elapsed time and minutes spent correcting the result. Higher effort earns its keep when the reduction in rework covers its cost. The API price list applies to another billing route. Work usage and cost

Give the spreadsheet a five-minute acceptance test

Recalculate one day from the source rows. Trace a cancelled order all the way to the formula for the total. Open the deliverable outside ChatGPT and change a number. Does the sheet recalculate, and does the accompanying explanation still hold? Record factual errors and the time it took to get a sendable report. A correct file that takes ten minutes to verify may earn a place in the weekly routine. An unexplained discrepancy still needs someone who can reconcile it.

Public web apps have their own hosting and access conditions. ChatGPT Sites

FAQ

Frequently asked questions

Do I need a separate subscription for ChatGPT Work?

OpenAI lists Work and Codex on Free, Go, Plus, Pro, Business, Edu and Enterprise. Features, model choices and allowances depend on the plan, client and workspace.

Does a local task run entirely offline?

Local execution can use approved resources on your computer. It does not promise on-device model inference. With sync enabled, OpenAI’s cloud can coordinate the work.

Does Work always use GPT-5.6?

Depending on account and client, the current model choices include GPT-6 Luna, GPT-6.1 Terra, GPT-6.1 Sol and GPT-6 Astra. Check the choices and permissions available in the intended workspace.

What does one assignment cost?

Model choice, context, tool calls and task length all affect usage. Work and Codex share usage within a ChatGPT plan. Measuring a representative test run gives a better estimate for a recurring task than applying API prices.

On this learning path

AI in your working dayStation 7 of 7

Last station

That is the whole path. The practice pack holds one exercise per station and the answers to every self-test.