Template · Putting it together
AI workflow map template
A task you do with AI every week is already a workflow, even if it only lives in your head. This template writes it down. It records each step, what goes in and out, which tool does the work, and where a person checks the result.
The most useful field is the review point. Review at the end catches errors after every later step has built on them. Review after the riskiest step catches them while they are still cheap to fix.
When to use it
- You run the same multi-step AI task every week or month.
- An error in an early step has reached the final output before.
- Someone else needs to run the workflow when you are away.
The template
Workflow name and output
What does the workflow produce, and for whom?
Steps
What are the steps, in order, each with an input and an output?
Tool for each step
Which tool or model class does each step use, and why?
Riskiest step
Where would an error do the most damage downstream?
Review point
Where does a person check the work, and what do they check?
Saved requests
Where are the requests and examples each step uses?
Owner and backup
Who runs it, and who can run it when they are away?
A filled example
A support team lead maps the weekly report that turns customer tickets into a summary of top issues for the product team.
- Workflow name and output
- Weekly top-issues report for the product team, sent every Monday.
- Steps
- 1. Export last week’s tickets (input: helpdesk; output: CSV). 2. Tag each ticket by issue (output: tagged CSV). 3. Count and summarize by tag (output: summary table). 4. Draft the report (output: one-page document).
- Tool for each step
- Step 1: helpdesk export. Step 2: a fast model, because the volume is high and I check a sample. Step 3: a spreadsheet, because counts must be exact. Step 4: a stronger model for the draft.
- Riskiest step
- Step 2. A wrong tag changes every count and every conclusion after it.
- Review point
- After step 2, I read twenty tagged tickets at random. If more than two are wrong, I fix the tagging rule and rerun the step.
- Saved requests
- The tagging request and the tag list are in the team folder, with five correctly tagged tickets as examples.
- Owner and backup
- Team lead runs it. The senior agent has run it twice and can cover.
What each field is for
Workflow name and output
Skill: Putting it together
Naming the final output and its reader sets the standard every step has to meet.
Steps
Skill: Putting it together
A step with a defined output can be checked. A step without one can only be trusted.
Tool for each step
Skill: Putting it together
Different steps need different tools. A spreadsheet counts more reliably than a model, and a fast model handles volume that a stronger one would make slow.
Riskiest step
Skill: Checking the work
Errors compound. Finding the step where they start tells you where review is worth the most.
Review point
Skill: Checking the work
A review point with a defined check and a defined threshold is repeatable. “Look it over” is not.
Saved requests
Skill: Putting it together
A saved request improves each time you fix it. A request rebuilt from memory loses every fix you made before.
Owner and backup
Skill: Getting real work done
A workflow only one person can run is a single point of failure. Naming a backup tests whether the map is complete.
Common mistakes
Splitting for its own sake
Steps exist so an intermediate output can be checked. A task with no useful checkpoint is often better as one well-specified request.
Putting all review at the end
By the final draft, a tagging error has shaped every count. Review where the error starts, not only where it shows.
Leaving the map out of date
When you change a step, change the map. A colleague following an old map will reproduce the problem you already fixed.