This user guide explains how to review and validate documents in Virtual Mailroom (VMR) by verifying automatically extracted data, correcting flagged items using lookup tools, and interpreting AI confidence and accuracy indicators.

• Reviewing Extracted Document Data
• Correcting Flagged Documents
• Understanding AI-Extracted Unstructured Data
• What's Next?

Reviewing Extracted Document Data

When a document is processed in VMR, the system automatically:

• Classifies the document type
• Extracts key information from unstructured data
• Attempts to map the document to the correct entity or account

Documents that require review appear in the Human‑in‑the‑Loop area.

During review, you can:

• View the source document alongside the extracted data
• Confirm the document has been classified correctly
• Validate key details such as:
  • Entity or fund
  • Document date and financial year
  • Investment code, member name, signature detection (where applicable)

This review step ensures the information extracted from unstructured data accurately reflects the source document before it is published.

Correcting Flagged Documents

Why documents are flagged
A document may be flagged for errors or concerns when one or more of the following conditions are identified during automated processing:

1. The AI confidence level for document classification is below the required threshold
2. The detected document type does not sufficiently match the document type classified   in VMR
3. Required information cannot be reliably extracted from unstructured data
4. Multiple potential matches exist for an entity, investment, or member
5. Key data points, such as document date or financial year, are missing or unclear

Flagged documents require user review and correction before they can be published.

Using lookup tools
VMR provides lookup tools to facilitate the resolution of flagged documents, enabling users to take corrective action, as outlined below, prior to publishing.   

1. Select the correct entity
2. Map the document to the correct investment holding code (APIR, ASX, Property, Unlisted) or member
3. Correct document dates or financial year values
4. Apply or adjust signature detection

Once all required fields are validated, the document moves to Ready status and can be published.

Understanding AI-Extracted Unstructured Data

What is unstructured data?

Unstructured data is information contained within a document that does not follow a predefined format and must be interpreted by AI.

Examples include:

• Names found in address blocks or headers
• Dates appearing in the document body
• Investment names or security codes
• Member names or signature indicators

During review, VMR displays how unstructured data has been interpreted and mapped, allowing you to clearly see what the AI has identified and confirm its accuracy.

Document type confidence logic

VMR assigns a confidence score to the document type classification by comparing the document type detected in the source document with the document type classified in VMR. This confidence score helps you quickly understand how closely the AI classification matches the source document.

VMR uses the following confidence logic and visual indicators:

High confidence (Green)

• Confidence ≥ 90%
• The document type on the source document exactly matches the document type classified in VMR
• Typically requires minimal or no review

Medium confidence (Orange)

• Confidence > 70% and < 90%
• The document type on the source document is similar or closely related to the document type classified in VMR
• Review is recommended before publishing

Low confidence (Red)

• Confidence < 50%
• The document type on the source document does not match the document type classified in VMR
• Manual review and correction is required before publishing

These indicators allow you to prioritise review effort and focus attention on documents that require validation.

How confidence is calculated

Overall confidence is calculated using a combination of exact match and fuzzy match rules applied to unstructured data extracted from the document.

Examples of matching logic used by VMR include:

• Exact matches (for example, fund name, trustee name, HIN or SRN)
• Partial matches
• Fuzzy matches where document text closely resembles known values

Each rule contributes a weighted score, which is then evaluated against defined confidence thresholds to determine the final confidence level displayed in VMR.

Confidence scoring supports accurate classification while ensuring that documents and data with ambiguity are clearly flagged for review.

Confidence levels support document review by helping users prioritise validation while maintaining accuracy and compliance.

Confidence indicators use colour‑based statuses to communicate classification accuracy and data validation outcomes. A green indicator signifies a strong match between the source document and VMR classification, including validated extracted data. An orange indicator indicates that the document should be reviewed to confirm classification or data accuracy, while a red indicator requires mandatory review and correction before publishing. Confidence scoring is derived from the analysis of unstructured data and applied matching logic.

What's Next?