--- title: What is a workflow description: What a workflow is, shown with a before-and-after example of one recurring job done without AI and with AI (which steps move to AI, which are shared, which stay human, and which new steps appear), plus the reasons the redesign is worth doing. Consult when someone is new to AI workflows, asks what a workflow is, or needs to see what changes when AI joins one. category: foundations updated: 2026-09-25 --- # What is a workflow **A workflow** is the set of steps a person performs, in a particular order and in a particular way, to get a bigger task done. Most work that looks like a single task is a workflow underneath. "Review the course evaluations" or "send the Friday report" names one job, but doing it means exporting, reading, sorting, writing, and deciding, one step after another. Seeing those steps is the starting point for working with AI, because AI rarely takes over a whole job at once. It takes over particular steps. The design question then becomes which steps AI does, which it does together with the person, and which stay entirely human. Asking a chatbot to handle a whole job in one prompt leaves the steps implicit and expects a finished result, which is the one-shot fallacy. Building an AI workflow makes the steps explicit and gives AI the ones it can do reliably, every time the job comes around. See: ai-workflow-redesign. Which mode a workflow runs in (chat, context, clockwork, or colleague) is a separate decision, made once the steps are known. An AI workflow can run in any of the four. See: ai-work-modes. ## Before and after: one workflow with and without AI The example used in teaching is reviewing over 5,000 open-ended course evaluation responses every quarter. As a to-do item, it is one task. Opened up, it is a workflow. Without AI, the job has six steps, all done by the person: export the responses, read them, save the interesting ones, sort them into themes, write a summary, and decide what to change. With AI, the same job has eight steps. The table shows who does each one. | # | Step | Without AI | With AI | |---|------|------------|---------| | 1 | Export the responses | You | You | | 2 | Make sure the responses are anonymous | Not a step | You (new step) | | 3 | Read the responses | You | AI | | 4 | Save the interesting responses | You | AI | | 5 | Sort them into themes | You | AI | | 6 | Spot-check the themes against the responses | Not a step | AI and you (new step) | | 7 | Write a summary | You | AI and you | | 8 | Decide what to change | You | You | Three things stand out in the comparison. Every original step survives. The job still involves exporting, reading, sorting, summarizing, and deciding; what changes is who does each step. Whatever time the redesign saves comes from the reading and sorting that moved to AI, not from dropping steps. The AI version is also two steps longer, because it has two steps the manual version never needed. Making sure the responses are anonymous before they go into an AI tool protects the people who wrote them. See: ai-diligence. Spot-checking the themes against the original responses before trusting them is how the person verifies work they no longer do by hand, and in this workflow responses the AI cannot place confidently are flagged for the person instead of being forced into a theme. See: ai-output-verification. Finally, the steps split three ways. Some are given away (reading, saving the interesting responses, sorting into themes), some are done together (spot-checking, writing the summary), and some are kept (exporting, anonymizing, deciding what to change). AI does the routine work, and the person keeps the thinking work. Deciding what to change in the course is the reason the workflow exists at all, and it stays human. See: productive-friction. In Danish-language teaching, the three owners are labeled "AI", "AI og dig", and "dig". ## Drawing your own before and after The same comparison works for any recurring job, and drawing it is the core of the Map phase. Start by naming the job the way it appears on a to-do list. Then list the steps it takes today without AI, in order, including the small ones that feel too obvious to mention, such as exporting a file or opening a spreadsheet. Next, mark each step with who should do it once AI is involved: AI, AI and you, or you. Finally, ask two questions the manual version never had to answer: what has to happen before information goes into the AI tool, and what has to happen before you trust what comes out. In the evaluation example, the answers became the two new steps. A more detailed decomposition names the workflow's trigger, inputs, decisions, outputs, tools, and frequency alongside the steps, and rates each step for how easy it is to verify and how risky it is to get wrong. See: ai-workflow-redesign. ## Why hand steps to AI The evaluation example shows four reasons to redesign a workflow with AI. **Before, not after.** The answer arrives while there is still time to act on it. The evaluation analysis moved from quarterly to weekly, so the course could be adjusted while it was still running. **Better, not just faster.** The saved time goes into depth. Time no longer spent reading went into following up with the people whose feedback was worth digging into. **What you never got to.** Work that had been dropped becomes possible. Analyzing the responses every week was impossible by hand. **Time for the thinking work.** The AI sorts, and the person decides. ## How a workflow gets built Building an AI workflow follows four phases. Map finds the job and breaks it into steps, which is where the before-and-after comparison is drawn. Test tries the AI version on real work. Integrate builds the tested version into everyday work. Compound makes the workflow better each time it runs. What each phase involves, and what it produces, is described in ai-workflow-redesign. ## Choosing which workflow to start with A good first workflow is structured (the same steps every time), repetitive (it comes around often, ideally every week), and easy to verify (the result can be checked quickly). Of the three, easy to verify matters most. See: choosing-a-first-workflow. ## Related pages - See: choosing-a-first-workflow — How to find and pick the first workflow to redesign, including the "week on repeat" exercise and the three criteria - See: ai-workflow-redesign — The four-phase method (Map, Test, Integrate, Compound) in detail, including the one-shot fallacy and the JIOPR delegation framework - See: ai-work-modes — How to choose the mode a workflow runs in once its steps are known - See: ai-output-verification — How to calibrate checks like the spot-check step to the stakes of the work - See: ai-diligence — Data handling decisions behind steps like anonymizing the responses