An automation demo is easy: feed a document to a model and print the answer. Production is different. Documents arrive in six formats, half of them scanned. Someone must confirm the output. Someone must fix it when the model is wrong.

The four parts that matter

  1. Intake — a single, validated entry point for every input.
  2. Extraction — deterministic parsing first, model inference second.
  3. Review — a queue where a human confirms or corrects, with the original document visible.
  4. Record — every decision stored with the input, the output and the reviewer.

Only one of those four is a model.

Confidence is a product decision

A pipeline should state when it is uncertain instead of guessing. We define a threshold, route low-confidence cases to review, and measure how often the queue is right.

Start with the workflow, not the model

If a process cannot be described on one page, automating it will multiply the confusion. We map it, remove the unnecessary steps, and only then write code.

That order is slower for a week and dramatically faster for a year.