How AI and Step-by-Step Thinking with GPT-4 Can Transform Financial Reviews
Accountants are increasingly turning to advanced technologies to enhance their analytical capabilities. One such technology is the use of large language models (LLMs), such as GPT-4, which have demonstrated potential in performing tasks traditionally reserved for human experts. A recent study provides compelling evidence of GPT-4's capabilities in financial statement analysis, highlighting its ability to predict future earnings without additional narrative or industry-specific context.
Key Research Insights
Comparison with Human Analysts
In the study, AI systems, specifically large language models like GPT-4, were better at predicting changes in a company’s earnings compared to human financial analysts. This is impressive because the AI did this just by looking at the numbers, without needing to understand the wider market or industry trends that human analysts usually consider. For accountants, this means that these AI tools can quickly analyze financial data and make accurate predictions, potentially saving time and reducing human error.
Chain-of-Thought (CoT) Prompting
Chain-of-Thought prompting is a technique where the AI is guided to think step-by-step, like a human would. For example, it might first look at changes in sales or expenses, then calculate important ratios like profit margins or debt levels, and finally, make a prediction about the company’s financial health. This approach greatly improved the AI’s accuracy because it systematically broke down the analysis process, making it thorough and logical. For accountants, using CoT could streamline complex analyses, ensuring nothing is overlooked.
Economic Insights
The AI models were not just crunching numbers; they could also provide insights by telling a story about the company’s future performance based on financial data. This means the AI could help accountants understand and explain what the numbers might imply about future profits or losses. By incorporating CoT prompting in your practice, you can leverage the depth of analysis it provides to enhance decision-making, offer more comprehensive client services, and maintain a competitive edge in financial consulting.
Chain-of-Thought Prompting Explained
The Chain-of-Thought (CoT) prompting approach is a method used with large language models to mimic human-like reasoning by structuring prompts that guide the model through a step-by-step analytical process. This technique was applied in the study to improve the model's ability to perform financial statement analysis.
How Chain-of-Thought Prompting Works
CoT prompting involves breaking down complex problems into simpler, logical steps, which the model then follows to reach a conclusion. This mimics the natural thought processes humans might use to solve problems, allowing the model to produce more structured and reasoned outputs.
Examples of CoT Prompting in Financial Analysis:
Analyzing Trends and Ratios:
Prompt: "Identify key trends in the following financial statement items: revenue, cost of goods sold, and net income. Calculate key financial ratios such as the current ratio, debt-to-equity ratio, and return on equity."
Model's Thought Process: The model first notes changes in revenue, cost of goods sold, and net income over the periods presented. It then calculates the requested ratios, noting any significant changes or deviations from industry norms.
Synthesizing Information:
Prompt: "Based on the trends and ratios identified, provide insights on the company's operational efficiency, liquidity, and financial leverage. Discuss how these might affect the company's earnings in the next fiscal year."
Model's Thought Process: The model synthesizes the trends and ratio outcomes to comment on the company’s efficiency (e.g., improved revenue but increasing costs suggest decreasing efficiency), liquidity (e.g., a decreasing current ratio might indicate liquidity issues), and leverage (e.g., a high debt-to-equity ratio might pose risks in debt servicing).
Forming Expectations About Future Earnings:
Prompt: "Considering the financial analysis completed, predict whether the company's earnings are likely to increase or decrease in the next year. Provide a rationale for your prediction based on operational efficiency, liquidity, and financial leverage."
Model's Thought Process: Based on the analyzed data, the model predicts an increase or decrease in earnings. For instance, if operational efficiency is declining, and liquidity issues are noted, the model might predict a decrease in earnings. The rationale explicitly ties these financial metrics to expected earnings performance.
Applying CoT in Your Practice
Custom Prompts for Clients: Develop custom CoT prompts tailored to the specific financial analysis needs of your clients. For example, for a client interested in understanding cash flow stability, the prompt could guide the model to analyze cash flow statements in detail, highlighting trends in operating, investing, and financing activities.
Interactive Analysis: Use CoT prompting in interactive tools where clients can input their financial data and receive a step-by-step analysis from the model, making complex financial assessments more accessible and understandable.
Training and Development: Train your staff on how to use and interpret CoT-based model outputs, enhancing their ability to provide nuanced financial advice.
LLMs in your practice, where to start
For accountants looking to incorporate AI technologies like GPT-4 into their practices, understanding the best method for integrating tax codes into these systems is crucial. Small practices may prefer the manual input method due to its simplicity and cost-effectiveness, allowing them to manually update and input the necessary legal documents as needed. This method provides a straightforward way to ensure that AI analyses are based on the most relevant and current information without the need for significant technical resources.
On the other hand, larger or more technologically advanced practices might opt for an automated API-based setup. Despite the higher initial costs and complexity, this approach offers significant long-term benefits in terms of efficiency, accuracy, and the ability to handle a larger volume of complex data seamlessly. Ultimately, the choice between these methods depends on the specific needs, resources, and strategic goals of the accounting practice, with each offering distinct advantages in the integration of cutting-edge AI tools into the demanding field of tax and legal compliance.
Integrating Tax Laws into AI Systems: Manual Input vs. Automated API Setup
For many small accounting firms, leveraging artificial intelligence (AI) such as GPT-4 for financial analysis requires incorporating relevant sections of tax laws into the system. There are two main approaches to achieve this: the manual approach and an automated API-based setup.
Here’s a comparison to help you decide which might be right for your practice.
Manual Input Approach:
This method requires accountants to manually select and input relevant sections of the tax law directly into the system before analysis. This might mean typing out or uploading digital copies of tax documents the AI needs to reference.
Automated API-Based Setup:
This more sophisticated approach connects your AI system directly to a continuously updated database of tax codes and rulings via an API (Application Programming Interface). This setup automatically feeds the latest legal information into your AI tools.
This method is exemplified by the innovative strategies employed by Harvey, a company that has revolutionised legal research for professionals. They worked with OpenAI to create a special program designed just for legal work. This program can sift through vast amounts of legal documents, much like searching through a huge, complex library quickly and effectively.
How Harvey Did It:
Starting Small and Growing Big: Harvey began by focusing on legal cases from Delaware and gradually included information from all over the U.S. This shows that starting with a manageable project and expanding as you get more comfortable can be a smart way to integrate AI.
Real-Time Data for Real-Time Decisions: They made sure their AI system always had the latest and most comprehensive data. This real-time updating is key to providing accurate and timely assistance, which is crucial not just in law but in any field that relies on staying up-to-date with regulations and information.
Making Lawyers’ Lives Easier: The AI tool that Harvey developed helps lawyers by pulling together important information quickly. This reduces the hours spent on looking up and cross-referencing documents, so lawyers can instead focus on building their cases or advising their clients more effectively.
Applying Harvey’s Strategy to Accounting:
Harvey’s story offers valuable lessons for accounting practices looking to adopt AI:
Tailored Tools for Specific Needs: Just as Harvey built a tool specifically for legal professionals, accounting firms can develop AI systems that are tailored to handle the intricate details of tax laws or financial reporting. This ensures that the services provided are not just general but specifically catered to the unique needs of the firm’s clients.
Grow as You Go: The idea of starting with a focused application and then expanding as needed can also work well for accounting. As the firm grows and handles more complex data or takes on clients from different areas, the AI system can adapt and expand, too.
Save Time on Routine Tasks: With AI taking over the more repetitive tasks, accountants can shift their focus to areas that add more value to their clients, like strategic financial planning or personalized consulting. This not only makes the firm more efficient but also enhances the quality of service provided to clients.
Steps for Using LLMs with Manually Provided Tax Law Sections
In scenarios where you need to manually identify relevant sections of the tax laws before using a large language model (LLM) like GPT-4 for analysis, the prompts you use would shift to include more specific referencing and validation steps. This process involves ensuring the model uses the correct legal and tax framework to analyze the data and make predictions:
Step 1 Manual Identification and Input:
You would first need to manually identify the relevant sections of the tax law that apply to the financial data or the tax scenario in question.
Once identified, input or upload these sections into the system where the LLM can access them. This might involve digitizing text if it’s not already in a usable digital format.
Step 2 Prompting for Specific Analysis:
After uploading the necessary tax code sections, your prompts to the LLM should specifically reference these sections to ensure the analysis adheres to the correct legal framework.
Example Prompt: "Given the tax code sections uploaded, specifically section [XYZ] related to capital gains tax, analyze the financial transactions list for potential capital gains events. Classify each transaction according to the rules outlined in the section provided."
Step 3 Validation of Model Outputs:
Even after providing specific prompts, it’s crucial to validate that the LLM’s outputs are consistent with the tax code sections provided.
Example Prompt: "Review the model’s classification of transactions under the capital gains tax rules from section [XYZ]. Confirm the accuracy of classifications and identify any discrepancies with the tax code."
Sample Prompts After Manual Tax Code Input:
Tax Liability Calculation: "Using the sections [XYZ] of the tax code provided, calculate the estimated tax liability for the following income statement items. Apply the tax rates and rules as per the uploaded sections."
Tax Planning Scenarios: "Based on the uploaded section [ABC] on tax deductions, simulate three tax planning scenarios for the revenue streams listed. Evaluate which scenario optimally reduces tax liability while complying with the tax code."
Compliance Check: "Cross-verify the deductions claimed in this tax return against section [DEF] uploaded from the tax code. Highlight any items that may not comply with the current legislation and suggest alternatives where possible."
Conclusion
The integration of LLMs like GPT-4 into accounting practices via CoT prompting represents a significant shift towards data-driven, AI-enhanced financial analysis. As these technologies continue to evolve, they promise not only to augment the existing capabilities of financial professionals but also to redefine what is possible in the realm of accounting and financial advisory services. Accountants are encouraged to explore these technologies, understanding their potential to bring about greater accuracy and efficiency in financial analysis.
Become a CIBA member and qualify to become an Accounting Officer and Independent Reviewer
Become a Successful Business Accountant with the Chartered Institute for Business Accountants
We help you become a successful business accountant by awarding top-quality designations that boost your relevance, credibility, and income potential within the finance department.
Here's how you can earn our prestigious designations:
Membership: Join our community.
Qualifications: Gain the necessary educational background.
Experience: Build your practical experience.
Competency Assessment: Prove your skills through our assessment.
By following these steps, you can achieve a designation that suits your career level and role. Here’s how you can get started:
Strategic Level: If you're a financial executive, aim for the Certified Financial Officer (SA) designation.
Management Level: If you're a financial manager, the Certified Business Accountant (SA) designation is for you.
Operational Level: If you're a financial accountant, pursue the Business Accountant (SA) designation.
Administration Level: If you're a financial administrator, go for the Chartered Financial Administrator designation.
Support Level: If you're a bookkeeper, the Chartered Bookkeeper designation is the perfect fit.
Practice Level: If you're an accountant in practice, aim for the Business Accountant in Practice (SA) designation.