Use this when

An AI pulls values from unstructured input, an email, a document, a form, and the workflow will act on those values.

The pattern

UNSTRUCTURED INPUT
      ↓
AI EXTRACTS values into fields
      ↓
RULES VALIDATE each field (format, range, match)
      ↓
VALID? ── Yes → workflow proceeds
        └─ No  → send to a person

The AI does the interpreting; deterministic rules check its work before anything depends on it.

Why it works

Extraction is interpretation, so it can be wrong, and a wrong value often looks clean and confident. Validation catches many of those errors cheaply: a total that is not a number, a date in the past, an order reference that matches no real order. It puts a deterministic guard between a probabilistic step and a real action.

This is the practical way to turn unstructured data into something the rest of the workflow can trust.

How to apply it

  1. Define the fields to extract and their expected form.
  2. For each field, write validation rules: format, range, and matches against known records.
  3. Let valid results flow on; send anything that fails to a person.
  4. Watch which fields fail most often; that points to input or extraction problems worth fixing.

Good fit

  • Invoices, forms, delivery notes, orders, any document-to-fields task.
  • Anywhere extracted values feed a consequential step.

Poor fit

  • Fields that cannot be meaningfully validated, where you must rely on review instead.

What this does not solve

Validation catches many extraction errors, not all of them. A plausible, well-formed, wrong value can pass. Combine it with a confidence threshold and, for consequential cases, human review.