Deterministic document extraction: why "AI reads your invoice" needs guardrails

Ask an LLM to summarise a contract and it does a reasonable job. Ask it to give you a compliance score you'll act on, and run the same document through it twice, and you may get two different scores. That inconsistency is fine for a summary and unacceptable for anything downstream of it.

Separate extraction from scoring

In DocuLens AI, the LLM's job stops at extracting structured fields — vendor, amount, date, line items. A schema-validated, deterministic layer then scores those fields against defined rules. The LLM's variability is contained to the part of the pipeline that can tolerate it.

Schema validation is not optional

If the extracted output doesn't match a defined schema, reject it and retry or flag it — don't let malformed output silently propagate into a compliance score someone will act on.

The rule of thumb

Let the LLM read. Let deterministic code decide. The moment a decision needs to be reproducible and explainable, it shouldn't be resting on a single LLM call's non-deterministic output.