Field notes
How to map your first AI agent workflow
Most workflow diagrams are beautiful and useless.
They capture how you wish the process worked. Not how it actually works. You draw the happy path. You leave out the exceptions. You skip the judgment calls because they are hard to put in a box. Then you try to automate a diagram that nobody actually follows.
This is not a flowchart problem. It is a clarity problem.
The diagram is not the map
A workflow map for an AI agent does not need to be visual. It needs to be honest. An agent follows your instructions exactly. It has no tolerance for ambiguity, no ability to infer what you meant, and no patience for undocumented edge cases.
The most useful thing you can produce before you build an agent workflow is not a diagram. It is a set of clear answers to five questions about one process.
Five questions that map any process
1. What triggers this process?
Something starts the work. An email arrives. A form gets submitted. A calendar event fires. A number crosses a threshold. Name the exact trigger and where it comes from. Not "when a customer contacts us." The specific signal and its source.
2. What information does the agent need at the start?
Before any decision or action, the agent needs inputs. What data does it need to begin? Where does that data live? What format is it in? What happens if it is missing?
This question surfaces the source of truth problem that stops most automations. The data you think you have in one place is often scattered across three tools, a spreadsheet, and someone's memory.
3. What decisions does this process require?
List every point where a human currently makes a judgment call. For each one: what information does the decision depend on? What are the possible outcomes? Is there a rule that covers the common cases? Is there an exception that breaks the rule?
If a decision is entirely tacit, you have found your bottleneck. Fix the decision first. An agent cannot inherit institutional knowledge it was never given.
4. What action does the agent take, and where?
An agent does something. It sends a message, updates a record, creates a task, routes a ticket, fires a webhook. Name the exact action, the target system, and what a successful execution looks like. Name what a failed execution looks like too.
If you cannot describe failure, you are not ready to automate. Knowing what success looks like is not enough.
5. What evidence does it leave behind?
Every automated action should produce a record. What gets logged? Where? Who can check it? What would tell you the agent made the wrong call?
This question matters more than most people think. An agent running without evidence is not automation. It is a process you have chosen not to watch.
What to do with the answers
Write them down for one process. Not your most complex process. The one that repeats most often and costs you the most attention each week.
Thirty minutes of honest writing is enough.
Read it back. If every answer is specific and complete, you are ready to build. If any answer is vague, contains "it depends," or references information you cannot locate, you have found the gap that would break the automation before the first run.
Fix the gap first. Then build.
This is the correct order. Not tool selection, not prompt engineering, not which AI model to use. First: can you answer five questions about the process you want to hand off?
If yes, you have your workflow map. If not, you have your next task.