Find One Work Task an AI Agent Can Take This Week
Run a 10-day audit to pick a recurring workflow an AI agent can run end to end, then automate it with before and after notes.
You are assembling the same file, summary, or draft again. The fastest way to take it off your plate is to assign an AI agent the whole recurring workflow, not another isolated draft.
Uber expanded ChatGPT Work from a 1,000-person pilot to more than 30,000 employees on October 22, 2026. In Uber's first pilot, 80 percent of participants were non-engineers, and they mostly used the tool to analyze data, spot business trends, and smooth out workflows.
ChatGPT Work lets a user state a desired result, after which the system collects context from connected apps, splits the job into steps, and can operate for hours to return a finished file or web app. Scheduled Tasks lets the agent handle repeat or event-triggered jobs by itself, for example summarizing customer feedback emails and sending the summary to the relevant Slack channels.
A finished workflow is easier to trust than a draft
Isolated drafts are easy to dismiss. A finished file, a routed summary, or a web app is harder to ignore. When an agent handles a complete job, you can compare the before and after without turning every task into a test.
You are also designing a repeatable process, checking its output, and deciding when a human should step in. That is the kind of work that survives a changing job description.
Pick the workflow before you pick the tool
Run a 10-day agent workflow audit with a teammate. Start with the work you already repeat, and find a job with a clear trigger, known inputs, and a result you can check.
- List five recurring multi-step tasks you repeat at least monthly. Include the trigger, inputs, output, and the person who needs it. Finish with a short table you can review in one pass.
- Mark which one an agent can run end to end. Look for a task where the agent can gather context, transform it, and deliver a finished file or web app without you approving every step. You are done when one row survives a skeptical review.
- Choose the task with the clearest output. A finished file, web app, or routed summary is easier to judge than a vague recommendation. The result should be something you can review quickly and send to the right person.
- Connect the agent to the apps where the work already lives. ChatGPT Work links Uber employees to Google Drive, Gmail, Google Calendar, and Slack.
- Automate the workflow with a schedule or trigger. The output should appear without you starting it each time.
- Record before and after notes for a full cycle. Note the time you spent, the errors you fixed, and the part of the prompt that did not work. Finish with a short memo you could hand to a teammate.
- Build a small library of prompts and checks for the task. Uber's rollout included an internal AI skills library, approved integrated apps, and dedicated training. The end state is a saved prompt, a list of required sources, and a quick check for missing context.
Avoid a task that needs human judgment at the end. If the final step is deciding what to do next, the agent will stall or produce something you must rewrite.
Judge it by whether it runs again
Do not measure success by how impressive the first output feels. Measure it by whether the workflow ran, whether the output was usable, and whether you would run it again. Uber's rollout will be judged by adoption rates and workflow outcomes rather than direct productivity quantification. Use that standard for your own test.
The signal is strongest in non-engineer knowledge work. More than 5 million people used Codex, a coding agent, weekly, and more than 1 million of those users, one in five, used it for work outside software development. People are already using agents for analysis, tracking, and coordination.
This week, start with the task that annoys you most but does not require a final decision. Run it, note the failures, and schedule the next run. Do it before the next recurring deadline.