Flo's Notes

AI in Cost Accounting

4 min read

Parts of this page were generated by AI.

Resources

AI in Cost Accounting: Capabilities and Pitfalls

🚀 Executive Summary

This lecture explored the intersection of Large Language Models (LLMs) and cost accounting. The central thesis is that while AI is a revolutionary tool for generating text, summaries, and code (like SQL queries), it is fundamentally unsuited for direct mathematical calculation or making autonomous accounting assumptions. The future of AI in finance lies in hybrid systems: using the LLM as an intuitive interface that writes queries for deterministic tools (like solvers or databases) rather than trusting the LLM to do the math itself.

⚠️ The Dangers of “Raw” LLMs in Accounting

1. The “Casino” Effect & Overconfidence

  • The Illusion of Competence: If an AI correctly answers an accounting query 9 out of 10 times, it builds false confidence. Users let their guard down for the 10th query, which the AI might hallucinate with the exact same level of absolute confidence.
  • The Slot Machine Analogy: A system that gives you the right answer most of the time but fails randomly is more dangerous in accounting than a system that always fails, because it tricks the user into skipping verification.

2. Hidden Assumptions in Cost Types

  • Context Dependency: Is electricity a fixed cost, variable cost, direct cost, or indirect cost? There is no universal truth; it depends entirely on a company’s specific historical treatment and business model.
  • The LLM Blindspot: An LLM will simply guess the classification based on general training data, entirely ignoring the company’s specific context. This leads to wildly varying results depending on the prompt or the AI’s temperature.

🛠️ How to Actually Implement AI in Cost Systems

To make a cost accounting system “AI-ready,” companies must separate the language interface from the calculation engine. LLMs must act as translators, not calculators.

1. The AI as an Interface (e.g., SAP Joule)

Modern ERP systems (like SAP) are building AI layers on top of their deterministic tools.

  • The Workflow: A user asks via a chat interface (like WhatsApp or a dashboard), “Which cost centers exceeded budget last quarter?”
  • The Translation: The LLM does not read the invoices and add them up. Instead, it translates the user’s intent into a deterministic query (e.g., an SQL SELECT statement).
  • The Execution: The ERP system executes the SQL query deterministically, ensuring 100% accuracy and reproducibility. The LLM then formats that accurate data into a readable summary.
  • Crucial Setting: When generating code or queries, the LLM’s temperature must be set to 0 to ensure strict, deterministic output without creative hallucinations.

2. Solving Complex Allocations (The Reciprocal Method)

The lecture examined a Module 3 task involving the reciprocal (iterative) method for allocating primary overhead between two service cost centers.

  • Raw LLM Failure: When asked to solve the problem directly, the LLM confidently produced the wrong numbers. It made structural errors, such as allocating property costs to maintenance when that wasn’t required.
  • The Correct Architecture:
    1. The LLM reads the problem and successfully sets up the system of linear equations.
    2. The LLM passes those equations to an external, deterministic mathematical solver (e.g., a Python script).
    3. The solver calculates the exact allocation rates.
    4. The result is perfect.

🔮 Structural Shifts in Enterprise Data

From Relational DBs to Knowledge Graphs

To optimize for AI, major software providers (like SAP, who recently filed patents for this) are shifting away from traditional relational databases (SQL) toward Knowledge Graphs.

  • Why? Knowledge graphs natively map the relationships between data points, allowing LLMs to traverse and retrieve interconnected financial data much faster and more reliably than writing complex, multi-join SQL queries.

The Auditing Problem (The Wirecard Example)

  • Training Data Bias: LLMs predict the most likely next token based on their training data.
  • The Blindspot: If an AI had audited Wirecard before the scandal, it would not have flagged the missing €2 billion. Why? Because the historical data (audited by EY) consistently reported the money was there. The LLM would simply assume continuity, making the exact same error as the human auditors.
  • Takeaway: AI auditing tools are currently limited by the historical validity of their training data. They cannot natively detect unprecedented fraud if the “ground truth” data they were trained on is fabricated.

Graph View

Backlinks