What is an AI workflow?

What separates a workflow from a prompt?

Four things. Miss one and you have an experiment.

• A named trigger. Every Monday at eight, or every time a form is submitted, or on request.
• Defined inputs. Which accounts, which pages, which spreadsheet, which criteria.
• An output that lands somewhere: a filled sheet, a CRM record, a drafted email in your outbox.
• A human checkpoint on anything consequential. Approve before it sends, before it deletes, before it writes to the system of record.

What does a real AI workflow look like end to end?

Weekly reporting, because everybody has it.

Trigger: Friday, 4 p.m. Inputs: six dashboards you are signed into, plus last week’s numbers. Steps: open each, pull the same three metrics, compare against last week, write two lines on anything that moved more than ten percent. Output: the sheet updated and a summary drafted to your team channel. Checkpoint: you read it before it posts.

In an agentic browser that whole thing is one instruction plus a schedule, and it runs in your logged-in accounts. Google Sheets and Slack are native integrations, so the writes go through tools rather than clicking. The patterns are at how to automate weekly reporting and generate a weekly metrics digest.

Why AI workflows fail

Three reasons, none of them about model quality.

The output has nowhere to land, so a person copies it by hand and then stops doing it. The trigger is somebody remembering, so it happens twice and never again. Or nobody wrote down what done correctly looks like, so nobody can tell whether the run worked.

The fix is process design. Give the output a destination, put the trigger on a schedule, write the acceptance check in one line. A companion covers all three: it writes into the sheet or the channel itself, a routine holds the schedule, and you read the result before it posts. The prompt half is at how to prompt an agentic browser.

Frequently asked questions