AI task spotlight | Edition No. 05: Document Analysis Summary

Every two weeks, we spotlight an AI task from Unit21's task library, something many compliance and fraud teams are configuring and running inside their workflows today.

‍

This edition covers Document Analysis Summary. The problem it solves is one every investigator recognizes: an alert arrives with a stack of attachments, and before the case can move, someone has to open each file, read it, and decide whether to trust it. A driver's license, a passport, a payslip, a bank statement, and a court record each need a different check. This task does all of that automatically and returns a single verdict.

‍

What document review actually requires

Documents are where investigations slow down. An EDD package, an onboarding file, a SAR-supporting court record, every one of these arrives as a stack of attachments that someone has to open, read, and interpret before the case can move forward. None of it is one task. A driver's license is read differently from a passport, and a payslip is read differently from a bank statement. Each document type has its own checks, fields, and ways of going wrong.

‍

So the baseline review is both repetitive and uneven. An investigator classifies each file, parses the machine-readable zone on a passport, checks whether the name on a payslip matches the entity on file, cross-references a bank statement, and summarizes a court record, all before reaching the one part that actually requires judgment: whether the documents hold up. Done carefully, that's a significant amount of time spent on mechanical work. Done under caseload pressure, it gets abbreviated, and the documents that should anchor a decision become the step everyone rushes.

The result is the same inconsistency that regulators expect compliance programs to eliminate. Some investigators read every attachment; some skim. Some catch the field-level mismatch; some never open the file.

‍

Introducing Unit21's AI task: Document Analysis Summary

Document Analysis Summary: What it does

Document Analysis Summary closes the gap between a document landing on an alert and an investigator knowing what it says and whether to trust it. The moment attachments arrive on an alert or case, the agent automatically classifies every file, routes each one to the analyzer built for it, and delivers a single summarized verdict grouped by entity, without the analyst opening a single file.

‍

Different documents require different checks, so the agent doesn't treat them the same way. Each file is sent to a specialized analyzer tuned for that document type, with a generic LLM fallback for everything else, so nothing on the alert goes unread.

‍

The agent automatically reviews:

• Every attachment on the alert or case is classified by document type before any analysis begins
• Identity documents, driver's licenses, passports, and national IDs, checked at the field level for verification at onboarding
• Fraud signals via MRZ parsing on passports and field-level mismatch checks across identity documents
• Payslips and proof-of-address documents, for income and address validation on lending or account reviews
• Bank statements, cross-referenced against the entity's data on file
• Court documents and other non-standard files, summarized through a generic LLM fallback so no attachment is left unreviewed

Document Analysis Summary: What the agent outputs

• A single summarized verdict per entity, consolidating every document into one read, so the investigator sees one answer instead of a stack of files
• A per-document breakdown of what each file is, what it says, and whether the fields check out, including MRZ and field-level mismatch findings on identity documents
• Multi-entity reconciliation across the alert, grouping documents to the right party so joint accounts and alerts with more than one named entity stay organized
• A surfaced result delivered directly on the alert or case. No downloading, no opening files, no manual review

Document Analysis Summary: Why this matters

Document review is exactly the kind of work that doesn't require human judgment until the very end. Classifying a file, parsing an MRZ, checking whether the name on a payslip matches the entity on file, summarizing a court document, these are mechanical, repeatable, and slow by hand. The judgment call, whether the verdict holds up, is the part that actually needs an investigator.

‍

The problem is that all of it is manual by default, and it scales badly. Every alert with attachments reveals the same bottleneck, and under caseload pressure the review is the first thing to get compressed. What follows is a compliance program that doesn't behave uniformly: the documents that should anchor a decision become the documents nobody fully read.

‍

Document Analysis Summary does that groundwork automatically. It classifies every attachment, routes each to the analyzer built for it, reconciles findings across multiple entities, and hands back one verdict before the investigator has opened anything. It works for identity verification at onboarding, fraud detection through MRZ and field-level checks, income and proof-of-address validation, bank statement cross-referencing, court document summarization for SAR support, and full EDD package review across multi-document sets.

‍

Because it's built with Unit21's Custom AI Agent tasks, it runs inside your existing workflow. Nothing new to learn. The documents get read, the findings are summarized and grouped by entity, and the judgment call, the part that actually requires a human, is all that's left.

‍

About the AI task spotlight series

The AI Task Spotlight runs every two weeks. Each edition covers one task from Unit21's library, covering what it does, how it works, and who it's for. If a task is solving a real problem for one team, it can probably solve the same problem for yours.

‍

Want to learn more? Sign up for a demo of our AI. Alternatively, stay informed of our AI by signing up for our next AI Task Spotlight.

‍