Identity fields were keyed manually from uploaded documents. Goal was faster structured capture without allowing model output to become an identity decision.

Automation had to reduce typing without weakening control.

Documents arrived as images. Fields needed canonical names, normalized dates, and compatibility with an existing onboarding contract. Model responses could be missing, malformed, or confidently wrong.

Model proposes. Code decides what enters system.

1. Multimodal extractionConstrained prompt requested only required identity fields from document image.
2. Canonical mappingApplication code mapped model output into existing domain fields.
3. NormalizationDates and English-name handling followed deterministic rules.
4. ValidationTests covered extraction and normalization behavior before downstream use.

AI handled narrow perception task. It did not approve identity or own onboarding contract.

Production AI is mostly boundary design.

Prompt quality mattered, but surrounding system carried reliability: input handling, schema mapping, validation, error behavior, and downstream compatibility. This pattern applies to invoices, applications, contracts, support attachments, and other document-heavy operations.

Manual field capture became assisted extraction.

Feature merged into onboarding path with dedicated test coverage. No claim is made that model independently verified a person, replaced compliance review, or achieved an unmeasured accuracy rate.