Where the path ends and judgment begins

When we sit down with a team to design a workflow, the first thing we draw is not a diagram of the steps. It is a line.

On one side goes everything that has a right answer a rule can find. On the other goes everything that needs somebody — or something — to read, weigh and decide.

Most of it is path

Teams are usually surprised by how much ends up on the rules side. Checking a total against an order. Looking up a customer by their account number. Routing a claim over a threshold to a senior handler. None of it needs intelligence; it needs to happen the same way every time.

Those steps are written as explicit logic. They are the parts of the process that should never surprise anybody.

Judgment gets one job

What is left is usually small, and usually the part that was eating the team's week: reading an email that could mean three things, matching a line description written by a supplier who does not use your product names, summarising a claim for the person who has to approve it.

That is where an AI step goes. It gets clear inputs, a narrow question and a shape its answer has to fit.

The gate is the design

The interesting decisions are about the gate between the two sides:

1. What can the judgment step decide on its own?
2. What does it have to hand to a person, and with what context?
3. What happens if nobody answers?

Get those three right and the rest of the workflow is plumbing. Get them wrong and no amount of model quality will save it.

The good news is that the answers are usually already in the room. The people doing the work know exactly which cases they would want to see. Our job is to write that down in a form a machine can follow.