Verification Guide

AISAY can check documents for anomalies, errors, and suspicious activity against pre-defined criteria. This guide covers common verification use cases including financial aid applications and document tampering detection.

Prerequisite: Read the Quick Start Guide first.
Tip: Verification tasks are particularly sensitive to the choice of models. Try both the Flagship and Local models to see which one works better for your use case. Local models tend to perform slightly better for verification.

What Is Verification?

Verification is about checking documents against rules you define. Instead of asking “what does this document say?”, you’re asking “is there anything suspicious about this document?”

Extraction	Analysis	Verification
“What is the account balance?”	“Is this person financially healthy?”	“Are there any suspicious transactions?”
“Who is the account holder?”	“Evaluate the quality of this report”	“Is this document tampered with?”

Example 1: Evaluating Suspicious Activity in Bank Statements

Scenario: You disburse financial aid to low-income households. Recipients submit monthly bank statements for verification. You need to flag items that warrant additional checks.

Tip: Put your overall context in the first field. The AI carries it through all subsequent fields.

Field 1 — Missing Information (Text):

I am a case officer for a Singapore Government Agency and one of my roles is to review applications for assistance to ensure that the cases are indeed needy and do not have any significant income sources.

To ensure this, applicants and their household members are required to submit their bank statements for one month. Bear this in mind for this entire process.

Check if the statements have any missing information, for example, the name or address, which could indicate that the applicant could be submitting someone else's bank statements.

Field 2 — Missing or Swapped Pages (Text):

Check if the statements have any missing, mixed or swapped pages. For example, someone could be trying to leave out a page that has a suspicious transaction. Check that all pages are present and in running order, and that there isn't anything suspicious in the page ordering.

Field 3 — Mixed Accounts (Text):

Check if there are more than one account in the bank statements, or if there are mixed accounts. We want to ensure the applicant submits ALL their statements, so if you detect more than one account being there, or a shortage of accounts, it could indicate that someone may be trying to hide something, or omitting information.

Field 4 — Large Transactions (Text):

Check if there is a presence of significant deposits, large cash withdrawals, large fund transfers. This could sometimes imply that this person has more sources or cash, or is making large spending on things that we do not know about.

Field 5 — Financial Difficulty (Text):

Check if there are signs of financial difficulty, this could mean the person is repeatedly drawing down on the account, or account balances at any time is dangerously low. We need to help these people if we see these signs.

Field 6 — Overall Recommendation (Text):

Based on all of the information that you have analysed, give me an overall conclusion and recommendation if I should approve the application, or seek a follow up review with the applicant.

Sample Output

When run on bank statements, AISAY produces structured results for each check. For example:

Missing Info: “The bank statements are missing crucial identifying information such as the applicant’s name and address. This is a significant omission that could indicate the statements do not belong to the applicant.”
Large Transactions: “There are several significant transactions present. On page 3, there is a deposit of 850.00 for ‘GOV GSTV’ and another deposit of 1,350.00 for ‘MSF-SSO Supplier Payment’. There are also large ATM cash withdrawals of 100.00 and 950.00.”
Overall Recommendation: “The application cannot be approved at this stage due to critical missing information. A follow-up review with the applicant is strongly recommended.”

Example 2: Detecting Document Tampering

Scenario: You receive licensing documents from applicants importing animals. You need to check if the certification or papers may have been forged.

Important: Use the Local model for this use case — it analyses both images and text, which is essential for detecting visual tampering.

Field 1 — Suspicious Text (Text):

I am a case officer for a Singapore Government Agency and one of my roles is to review import documents that applicants submit when trying to bring in their pets into Singapore. This typically requires them to get documents from overseas authorities or overseas parties.

I am concerned that some people may be tampering with their documents and forging it, and I need you be an expert detective here, to check all the documents for a list of criteria. Bear this in mind for this entire process.

For this field, check if the document has any signs of spelling mistakes, grammatical errors, unusual/ amateur phrasing or weird language that could give us any suspicion. List all the occurrences you find and use your own judgement to differentiate genuine mistakes. When in doubt, just raise the issue.

Example Response:
Occurrence 1: Short Title
Occurrence 1: Description and Reason

Occurrence 2: Short Title
Occurrence 2: Description and Reason

Field 2 — Suspicious Layout (Text):

For this field, check if the document has any signs of misaligned text, misaligned columns, misaligned logos, or weird layouts, which could give us cause for suspicion. Also check if there are any signs of document tampering, if something looks like it was cut away, or cut and pasted on this document.

Field 3 — Logical Checks (Text):

For this field, check if there are any logical errors that you observe in the document that could be suspicious. For example, if the company name is Disney and the website is called SUSPICIOUSURL.COM then it could be cause for concern.

If there are any pieces of information that do not make sense, or are inconsistent, or cause for suspicion please raise this here.

Field 4 — Summary and Recommendation (Text):

Based on all the information you have seen, provide an overall summary and recommendation on what I should do.

Tip: If responses come as one large block of text, add an “Example Response” section at the end of each prompt showing the format you want (with line breaks between occurrences).

Tips for Better Verification

Be explicit about your criteria — List exactly what counts as suspicious or non-compliant
Explain why each check matters — This helps the AI weight its findings appropriately
Use conditional logic — “If X is present, then check Y”
Request structured output — Ask for findings in a numbered format with reasons
Try both models — Local models tend to catch more visual anomalies; Flagship models are better at logical analysis of text
Keep humans in the loop — Verification results should inform human decision-making, not replace it. Use AISAY to surface issues that warrant a closer look.