--- title: Choosing a first workflow description: How to pick the first workflow to redesign with AI. Covers two ways to list candidates (the "week on repeat" exercise and a full working day of involving AI in everything), the three criteria that cross them off (structured, repetitive, easy to verify, with easy to verify weighing most), the two questions that remove more, and how to make the final pick. Consult when someone is choosing where to start, or asks whether a particular job is a good fit for AI. category: foundations updated: 2026-09-25 --- # Choosing a first workflow **Choosing a first workflow** is the decision about which recurring job to redesign with AI before any building starts. The choice shapes the whole experience. A job that is vague, rare, or hard to check tends to produce a disappointing first result, and the disappointment gets blamed on AI rather than on the choice. A job that fits the criteria below produces a result the person can check, use, and improve. The method has three moves: list candidates from the person's own week, cross off the ones that fail the criteria, and pick one from what remains. ## Listing candidates Candidates come from the person's own week, not from a list of things AI can do. The exercise: think of a typical week and imagine having to live that week on repeat, forever, without being able to change anything. Which tasks would you most want to get rid of before the endless loop starts? Writing them down produces a first shortlist. A real first shortlist from teaching had three items: a Friday report made by copying numbers from a spreadsheet into an email, over 5,000 open-ended course evaluation responses to read every quarter (the workflow broken down step by step in what-is-a-workflow), and meal planning, where the meals are the enjoyable part and the planning is not. A complementary way to build the list is to spend an entire working day collaborating with AI on whatever work naturally arises. Rather than choosing candidates in advance, the person works normally but involves AI at every step: asking for ideas, requesting proofreading, delegating research, seeking feedback. The day maps the practical boundary between where AI adds value and where it falls short for that specific person's work, so the list rests on experience rather than theory. The candidates it produces go through the same criteria as any other. ## The three criteria Each candidate is checked against three criteria, and any candidate that fails one is crossed off. **Structured.** The job follows the same steps every time, so the steps can be written down and handed over. **Repetitive.** The job comes around often, ideally every week. Frequency is what makes the setup effort pay back, and every run is another chance to improve the workflow. **Easy to verify.** The person can see quickly whether the result is right. This is the most important of the three. When checking the result means redoing the work, nothing is saved; when the result goes unchecked, mistakes travel on into decisions and documents. A result that can be checked in minutes, such as figures compared against their source or a summary compared against a sample of the originals, is what makes it safe to hand a step to AI at all. See: ai-output-verification. In Danish-language teaching, the three criteria are "struktureret", "tilbagevendende", and "let at tjekke". ## Two questions that cross off more Two further questions apply before the final pick. The first is whether the work should exist at all: automating a report nobody reads makes it harder to question later, and deleting it is the better redesign. The second is whether an existing tool already handles most of it: if a standard product covers 80 to 90 percent of the job, a custom AI workflow adds maintenance cost for little gain. See: ai-productivity-traps. ## Making the pick From what remains, the person picks one job: the one they would most gladly hand off, judged by the effort it costs rather than by how exciting it would be to automate, and with low-to-medium stakes if one run goes wrong. Then they start with one small piece of it. In the teaching example, the first job was the Friday report, and the first piece was narrower still. The AI read the training-completion figures and drafted the email to the manager; the person checked the figures and reviewed the email before sending it. ## Good fit or poor fit Jobs that make good first workflows tend to be recurring operational work built from the same inputs each time: weekly status updates, customer meeting preparation, campaign reporting, monthly forecasting, and support ticket triage. Jobs that make poor first workflows are usually too vague, too strategic, or too irregular. A vague job, such as "help with marketing", has no fixed steps, so it fails structured. A strategic job, such as "develop a product strategy", is hard to verify and is the thinking work the person should keep. An irregular job, such as an annual planning exercise, fails repetitive. These jobs can still get one-off help from AI in chat; they are poor candidates for a first workflow. See: ai-work-modes. ## The three criteria and the five-point filter The three criteria screen whole jobs and are quick to apply to a list. Once a workflow is chosen and broken into steps, each step gets a stricter check, the five-point filter: verifiable, step-wise and bounded, recurring and painful, inputs and outputs known, and a human in the loop. A workflow can pass the three criteria and still contain steps that should stay human, as the final decision step does in the evaluation example. See: what-is-a-workflow and agent-use-case-evaluation. ## Related pages - See: what-is-a-workflow — What a workflow is, with a before-and-after example of which steps move to AI - See: agent-use-case-evaluation — The five-point filter for rating the steps inside a chosen workflow - See: ai-productivity-traps — The build-before-delete trap and build-versus-buy distortion behind the two extra questions - See: ai-workflow-redesign — The four-phase method that follows the pick: Map, Test, Integrate, Compound - See: ai-output-verification — How to check AI output, and why easy verification makes a job a safe first pick