Validation and Review

AI extraction must use a human-in-the-loop model. The user remains responsible for confirming business data before it is saved.

Review principles

Level	Suggested threshold	Behavior
High	`>= 0.90`	Can be auto-filled without warning.
Medium	`0.70 - 0.89`	Auto-filled but highlighted for review.
Low	`< 0.70`	Manual review required.

Auto-approval should be limited to low-risk fields and mature extraction profiles.

Field type	Auto-approval recommendation
Document language	Usually safe.
Provider name	Safe only if matched to an existing provider.
Monetary amount	Require review unless matched against known total.
Contract dates	Require review.
Fiscal code, VAT, IBAN	Validate deterministically and require review when confidence is not high.
Tax regime	Require review.
Party role	Require review when multiple parties appear.

If document classification fails, ask the user to select the document type.
If OCR quality is poor, show a manual-entry fallback.
If mandatory fields are missing, keep the draft but block final save.
If two existing entities match the same extracted party, require user selection.
If the model returns invalid JSON, retry with a stricter schema prompt or fallback model.

Edge case	Handling
Missing data	Leave field empty and mark as required when needed.
Ambiguous date	Show extracted text and require user confirmation.
Multiple tenants	Create multiple party drafts and roles.
Multiple properties in one document	Ask user to select or split into separate records.
Bank account without holder	Save as draft only until holder is confirmed.
Scanned low-quality document	Run OCR enhancement or request a clearer upload.

Each saved record created from AI-assisted extraction should keep a trace to: