In 30 seconds

Some software follows rules. The same input always produces the same result. Some software interprets, and the result can vary and can be wrong.

Good AI workflows combine both: fixed rules where consistency matters, interpretation where judgment helps.

The simple model

A deterministic rule might say:

IF invoice total is above €10,000
THEN require manager approval

Run it a thousand times with the same invoice and you get the same answer a thousand times. This makes it easy to test, explain and trust.

An AI model might instead read an email and decide whether the customer sounds dissatisfied. That requires interpretation. Ask it twice and you might get two slightly different answers. This is what "probabilistic" means in practice: the output is a likely answer, not a guaranteed one.

Why the difference matters

If you treat a probabilistic step as though it were deterministic, you will design a fragile workflow.

You might assume the same document always produces the same extracted field, or that an approval message always uses the same words. When the model varies, the workflow breaks in ways that are hard to reproduce.

Once you know a step is probabilistic, you design for it: add validation, constrain the output, or put a person in the loop where consistency matters.

A small example

A workflow reads delivery notes and pulls out the quantity received.

For clean, printed notes the model is reliable. For a handwritten note with a smudge, it may read 12 as 72. A deterministic check ("does the quantity match the order within a tolerance?") catches the surprising cases without pretending the model is perfect.

The AI does the interpreting. The rule does the guarding.

What this does not solve

Knowing a step is probabilistic does not tell you how often it will be wrong, or how bad a wrong answer is. You measure that when you build a test set and evaluate real cases.

Good enough to move on when

  • You can label each step as rule-based or interpretation.
  • You know which steps must always produce the same result.
  • You have a plan to guard the interpretation steps that matter.