The Power of the Prompt: How Gen AI is Transforming Audit and Accounting 

Gen AI tools, like Caseware’s AI digital assistant, can make accountants and auditors more efficient by reducing the amount of mundane work they need to perform. But the quality of an AI’s response depends upon the quality of the question. If you want an accurate, useful response, you need to craft your AI prompts carefully.  

A guide for creating effective prompts: The STAR method  

Following a prompt writing methodology is an important step in generating better AI responses. A writing methodology is a structured approach used to craft clear, precise prompts. You set the context for your query, define the task the AI is to perform, specify your desired output format and once you receive a response you refine your prompt to get the exact output you want. 

One popular prompt engineering approach is the STAR Methodology. STAR stands for:  

• Situation – Set the context for your prompt by describing the accounting scenario you want the AI to address. For example, you could tell the AI your precise role, your goal and the accounting standards you need to follow. 

• Task – Clearly define what you want the AI to do for you.  

• Appearance – Tell the AI what you want the output to look like. You can ask for a specific tone of voice, response length or format, such as a table or chart. 

• Refine – Refine your initial prompt as many times as you need to get a response that meets your exact needs. 

Situation 

Situational context helps minimize the risk of an incorrect response by reducing ambiguity. Context gives your large language model a frame of reference which allows it to produce more relevant answers. 

Task 

Task is the area most AI users will be familiar with. It’s what you’re asking the AI to do. The more details you can provide, the better the response will be. You can also specify what knowledge bases you would like the AI to use during the task phase.  

A new gen AI approach, called Retrieval Augmented Generation (RAG) pairs AI with reliable, high-quality data sources, so the AI can pull from curated, authoritative knowledge bases, rather than just relying on its large language model for information. Your firm’s RAG could include your best methodology documents and your gold-standard memos, giving you a firm-specific model that outputs a consistent look and feel and gives you a competitive differentiator. 

Appearance 

When building your prompt, you need to specify what appearance you’d like the AI’s response to take. Some questions to ask yourself could include: 

• Do you need it to be a particular word count?  

• Do you want British English or U.S. English? 

• Do you want a checklist? 

• Do you want a spreadsheet table?  

• Do you want the response in a particular order, so it matches perfectly to a long-form checklist? 

The more you control your output in specific terms, the more efficiency you’ll achieve. 

Knowing where your AI’s answers come from is essential for fact-checking. Despite technology improvements, manual verification of facts is necessary to ensure you’re presenting reliable information.  

There are several ways to find this out. First, you can ask your AI for citations. In Caseware AiDA, if you’re asking questions based on a PDF, you’ll get clickable footnote references that link directly to the highlighted source within the document. This feature makes fact-checking much faster and more accurate.  

Another way to check your sources is to tell your AI which authoritative body of knowledge you want it to use and to provide links for where things are. This way, you can click through and double-check your information easily.  

You can also boost the transparency of your AI’s responses through “chain-of-thought reasoning.” This means asking your AI to document each step it takes to generate a response. You could ask the AI a question like you normally would and then add the instruction, “Explain your reasoning step by step.” Breaking down a complex problem into a series of intermediate steps allows the AI to reason through each step sequentially.    

Refine 

Refining an AI’s response involves reviewing it for accuracy and clarity and then editing as necessary. This could include shortening or lengthening a response, as well as copy editing.  

Some key points to consider when refining your response include: 

• Clarity – Make sure the AI’s response is phrased in straightforward, unambiguous language. Jargon or complex terms could confuse your audience.  

• Relevance to the task – Your refined prompt should directly relate to your task and its objectives. Remove any instructions that don’t contribute to your goal.  

• Completeness and accuracy – When you refine your prompt, be sure it’s complete and accurate. The refined prompt should also include any necessary caveats or considerations. 

• Iterative Refinement – You’ll seldom get the precise response you want on your first prompt. You can start with a draft prompt and then review and revise it multiple times to enhance it. Seeking colleague feedback can help you identify any areas in the AI response that might be lacking. 

Build a prompt library for consistency and quality 

Some prompts can be long and contain tasks that can be applied to other accounting-related prompts. This is why a prompt library with all your template prompts can be an incredibly useful resource for your team. By housing expert-crafted prompts in one place, your staff can use high-quality prompts to tackle specific tasks consistently. For instance, you might have your firm’s asset purchase agreement expert draft a detailed prompt for generating agreement analyses, saving time and ensuring consistency. 

Prompt libraries are particularly beneficial for junior staff, who can use them to analyze complex documents more accurately and avoid common errors, like missing critical contract clauses.  

Conclusion 

By adopting these prompt writing practices, firms can use AI more effectively, improving efficiency and accuracy. Keep in mind that AI is a complement to accountants and not a replacement. Ethics codes require accountants to take responsibility for the output of an AI. Always apply your professional skepticism to an AI’s responses to ensure you’re getting an accurate result.