Use the Model for the 10% That Needs One

In a dependable AI automation, most of the pipeline can be ordinary code.

Fetching an email, looking up a purchase order, writing to the ERP, and posting an approval are all jobs with defined inputs and outputs. A model might help interpret an unfamiliar attachment or classify a request in the middle. It doesn’t need to run the whole process.

The 10% in the title is a rule of thumb. The useful exercise is finding the smallest part of the workflow that needs a model at all.

Why I prefer this split

Failures are easier to trace when each step has a clear job. A failed API request gives you a status code to inspect. A classification error gives you a specific model input and output to review. A giant prompt that handles the entire workflow makes those failures harder to separate.

Costs are easier to understand, too. Code still costs money to run and maintain, but predictable steps don’t need a model call each time they execute. You can spend the model budget on the work that benefits from it.

The split also gives you a testing strategy. Use ordinary tests for the deterministic steps and a fixed set of real evaluation cases for the model. Then check the complete workflow, including what happens when a step fails.

An automation that behaves consistently is easier to trust. You can decide which steps may run unattended and which need someone to review the result.

Work through the steps

For each step, ask whether a script could handle it reliably. If it can, use a script. Keep the model’s input, output, and permissions narrow enough that you can test it or replace it without rewriting the rest of the system.