The concept
A business workflow is a sequence of steps that turns an input, such as an invoice, a support ticket or a job application, into an outcome, such as a payment, a resolution or an interview. Most steps in most workflows are deterministic: look up a record, apply a rule, send a message. A smaller number are judgment steps: read this, decide what it is, extract what matters, draft a response. Language models are good at judgment steps over unstructured text, and that is precisely where they belong in a workflow. Everything else should stay deterministic.
Integration, then, is not about replacing a workflow with a model. It is about identifying the judgment steps, wrapping each in a well-defined interface with inputs, outputs and confidence, and connecting it to the deterministic steps around it with the same discipline you would apply to any external service that is sometimes wrong.
Why it matters
Teams that treat a model as a magic box that handles the whole process build systems that are impossible to test, debug or trust. Teams that treat it as one component among many, with a narrow job and a measurable error rate, build systems that work in production. The difference is architectural, not technological. The same model produces reliable outcomes in the second design and unreliable ones in the first.
The judgment-step test
A step is a candidate for a model when most of the following hold:
- The input is unstructured or semi-structured text, speech transcript or document content.
- A trained person could do the step by reading, without looking anything up that the system could not provide.
- The output can be expressed as a small structure: a label, a few fields, a short draft.
- Errors can be detected, routed to a person, or corrected downstream.
- The step is frequent enough that automating it matters.
A step is not a candidate when it requires authoritative lookup (use a database), precise arithmetic (use code), irreversible action (use a human or a verified rule), or when the cost of a wrong answer is high and undetectable.
Worked example
An accounts payable workflow: an invoice arrives by email; someone reads it, identifies the vendor, extracts the amount, date, invoice number and line items, matches it to a purchase order, codes it to accounts, routes it for approval, and schedules payment. Judgment steps: reading the invoice to extract fields and identifying the vendor from messy sender data. Deterministic steps: matching to the purchase order by number, applying approval rules by amount, scheduling payment. The model handles extraction and vendor identification, each with a confidence score. Low-confidence extractions route to a person. Everything after extraction is code.
Common mistakes
- Asking a model to do the arithmetic or the lookup instead of the reading.
- Letting a model take an irreversible action directly.
- Integrating a model into a step that was never clearly defined in the first place.
Action Step: Choose a real workflow you know well with at least six steps. Write each step on one line. Mark each as deterministic or judgment. For every judgment step, apply the five-point test and record the result. You will carry this workflow through the whole course.