Introduction
Presented at AAAI-92, this research explores the intersection of artificial intelligence and financial auditing. It introduces SAVILE, a system designed to assist auditors in analyzing the complex "data processing machinery" of a corporation rather than merely verifying final numerical balances. By modeling the underlying systems that produce financial data, SAVILE enables a more rigorous evaluation of internal controls and audit efficiency.
The Challenge of Modern Auditing
Auditors are responsible for certifying the fairness of financial statements. However, these statements are the culmination of thousands of individual transactions processed through opaque, multi-departmental systems. A minor procedural error in one area can propagate through the system, leading to significant, non-obvious impacts on the final balance sheet. Traditional automated tools often focus on microeconomic variables; SAVILE addresses the "plumbing" of data processing—the policies, personnel, and workflows that constitute the accounting system.
Methodology: The SAVILE Framework
The core of SAVILE is a domain-specific ontology and a modeling language called SPLAT (SAVILE Programming Language Attributed to TICOM). The system defines four primary concepts:
- Records: Data objects such as invoices, checks, and purchase orders.
- Repositories: Storage locations like databases, folders, or physical in-boxes.
- Actions: Primitive operations including
copy,compute, orcredit/debit. - Activities: Sequences of actions performed by a specific group (e.g., "Cash Disbursements").
SAVILE utilizes Petri Net semantics to represent accounting workflows. Records act as "tokens" that flow through the system. When an activity receives all required input records, it "fires," executes its internal actions, and moves the resulting records to output repositories.
A key innovation is the use of a single "correct" system model that is transformed into different representations for specific tasks. For instance, Causal Networks are used to trace how "corrupted" records propagate to evaluate internal controls, while Dependency Graphs facilitate "path sensitization"—a technique borrowed from VLSI (Very Large Scale Integration) circuit testing—to determine which records must be sampled to prove a specific failure occurred.
Case Study: The Purchases and Payables System
To evaluate the system, the authors utilized a detailed case study of a representative "Purchases and Payables" (P&P1) system. This model represents the standard corporate workflow for ordering goods, receiving them, and processing vendor payments.
One of SAVILE's most powerful features is Automatic Failure Enumeration. By describing only the correct behavior, the user allows SAVILE to automatically generate potential "activity failures" (mistakes or omissions) by systematically substituting actions with their known failure modes.
Key Results and Analytical Power
Using the P&P1 case study, SAVILE demonstrated significant analytical capabilities:
- Failure Filtering: Of 40 possible system failures identified, SAVILE determined that only 28 were "audit significant"—meaning they could materially affect the financial statements.
- Control Evaluation: The system identified that 14 of these 28 significant failures would be caught by existing internal controls (such as matching an invoice to a receiving report), highlighting 14 undetected risks.
- Test Optimization: To address the remaining risks, SAVILE initially proposed 45 substantive tests. By analyzing dependencies, it merged these into just two complex tests, drastically reducing the manual labor required for the audit.
Conclusion and Future Directions
SAVILE demonstrates that while accounting systems are complex, they are composed of conceptually simple, interacting elements suitable for model-based AI. Future work aims to reduce the manual burden of initial model acquisition through graphical interfaces and to integrate knowledge-based risk assessment to account for the varying likelihood of manual versus automated failures. As a "progress report," this research lays the groundwork for evolving representations that can handle increasingly complex accounting tasks.
Summary updated