How to map an AI workflow: deterministic, fuzzy, irreversible
A simple exercise for teams building AI products: label every step in a workflow to decide what the model, the software and the user should each handle.
Before you pick a model or design a chat interface, take your most important workflow and give every step one of three labels. It’s a simple exercise, but it forces decisions that are easy to skip when you’re focused on the demo.
Why this matters
Imagine a workflow with ten steps, each with a 95% chance of working, where all ten have to succeed. If the steps are independent, the whole workflow succeeds only about 60% of the time.
That’s a simplified calculation, not a measured failure rate. Real systems have validation and retries. It still shows why model accuracy alone doesn’t tell you whether the product works.
The three labels
Deterministic. Clear rules, and the same input gives the same result. Validating required fields, checking permissions, preventing duplicate actions. Use conventional software for these.
Fuzzy. The step needs interpretation. Reading an email, classifying a request, drafting a reply. This is where AI earns its place.
Irreversible. The step has consequences that are hard to undo. Sending an external email, issuing a payment, committing a booking. These need explicit permissions, verification and, where appropriate, human approval.
An example: handling refund requests
- Read the customer’s email and work out what they’re asking for: fuzzy
- Find the order and check it against the refund policy: deterministic
- Draft a reply: fuzzy
- Issue the refund and send the reply: irreversible
The labels start to shape the interface:
- Step 1 appears as structured fields the agent can fill in and the user can correct, not as a paragraph explaining what the model found.
- Step 2 looks different from step 1, because a checked fact and an interpretation shouldn’t carry the same visual weight.
- Only step 4 asks for confirmation. If people have to approve everything, they eventually stop paying attention.
Three questions for each step
- What happens when this step is wrong?
- Who notices, and how?
- Can it be undone?
If you can’t answer those, you’ve found your next design problem.
The longer version
I wrote more about this, with examples from Routine Labs and Zauber, in UX Collective: Everyone is building AI agents. Who’s designing the stop button?