Companies usually have more ideas for AI than capacity to try them. The most useful early decision is not which tool to buy; it is which one workflow to automate first. A good first workflow teaches you quickly, carries limited risk, and produces evidence you can use to decide what comes next. This guide gives you criteria and a simple method to choose.
Start simpler than you think
Anthropic's guide to building effective agents recommends finding the simplest solution possible and adding complexity only when simpler solutions fall short. It separates workflows, where models and tools follow predefined code paths, from agents, where the model directs its own process. In our experience, a well-defined workflow is often enough for a first project, and agents fit better when the steps are hard to predict. Choosing a first workflow with predictable steps keeps your early learning fast and your risk low.
Seven criteria for a good first workflow
These are our practical criteria, not a standard.
- It happens often. Frequent work repays the effort of automation and gives you many examples to test.
- Inputs and outputs are clear. You can say what goes in, what should come out, and what a good result looks like.
- You can measure a baseline. You know today's time, volume, error rate, or cost, so you can show improvement.
- An error is affordable. A mistake can be caught and corrected without serious harm.
- The data is accessible and permitted. You can use the data legitimately, and you know where it lives.
- A person can review the result. There is a natural point for human approval.
- Someone owns the outcome. A named person cares about the result and will use it.
A simple scoring method
List three to five candidate workflows. Score each criterion from 1 (weak) to 3 (strong), add the points, and discuss the top two. Do not treat the score as the decision; use it to surface disagreements. If a candidate scores low on "an error is affordable" or "data is permitted," it is not a first project, however attractive it looks.
Good first candidates
- Classifying and routing incoming messages, with a person reviewing urgent cases.
- Drafting responses or documents from a template and reference material, with review.
- Extracting structured data from documents, with spot checks.
- Preparing a recurring report from several sources.
- Summarizing long threads or meetings into actions.
Workflows to avoid first
- Anything where an error is expensive or hard to reverse, such as financial or clinical decisions without review.
- Work that depends on sensitive or personal data you do not yet have a policy for.
- Processes nobody can describe, or that change every week.
- Fully autonomous workflows with broad permissions. OWASP lists "excessive agency" among the top risks for LLM applications, so grant only the access a workflow needs.
What to measure
Record the baseline before you build: time per item, volume, error rate, rework, and the effort of the people involved. After a short cycle, compare quality, time, manual effort, failure modes, and what the users think. Write down what did not work as well as what did.
Run it as a bounded cycle
Define the objective, the scope, the review points, and the success criteria in writing. Build a reviewable prototype with representative data, test it, and end with a decision: continue, revise, industrialize, or stop. That is the structure of our AI Iteration Lab, a monthly cycle designed to prove one opportunity before a larger commitment. For the quality side of working with AI, read From Prompts to Method.