Fieldwork & evidence analysis

Fieldwork is where the audit plan materializes into evidence. This phase has historically been dominated by manual inspection. However, the emergence of multimodal language models represents a paradigm shift, allowing us to automate the processing of complex documents at scale and transform how we approach substantive testing.

The core challenge of fieldwork is extracting structured data from unstructured sources—invoices, contracts, and bank statements. Multimodal models excel here because they interpret information by combining text with visual layout, just as a human would. This enables sophisticated analysis of scanned documents.

Consider the classic three-way match test. A prompt can operationalize this entire procedure.

You are an audit associate performing tests of detail. Given the attached scanned invoice (INV-123), purchase order (PO-456), and goods received note (GRN-789), perform the following:

1. Extract Vendor Name, Invoice Date, Total Amount, and Quantity for 'Product X' from all three documents.
2. Compare the extracted fields across the documents.
3. Present the findings in a table and explicitly state whether the details match or if there are exceptions. Think step-by-step as you perform the comparison.

For more complex tasks, such as compliance with ASC 842, few-shot prompting is an exceptionally powerful technique. By providing the model with a clear example of the input and desired output, you can train it "in-context" to replicate that logic across a large portfolio of documents.

Here is how you would structure that prompt:

Your task is to extract key data points from lease agreements for ASC 842 analysis.

Example 1 Input: [Text snippet from Lease A] 'This Lease commences on October 1, 2024, for a term of five (5) years. Monthly rent is $5,000.'
Example 1 Output: {
  "Commencement_Date": "2024-10-01", 
  "Term_Months": 60, 
  "Monthly_Payment": 5000, 
  "Renewal_Option": "Not Stated"
}

Now, using this format, process the attached 10 lease agreements and output the results as a single JSON object.

This approach converts hours of manual extraction into a single, automated task, freeing up auditor time for higher-value analysis of the results.