“What is the capital of Georgia?”
If you said:
- “Atlanta”, then congratulations, you are well versed with the 50 US states.
- “Tbilisi”, then well done for recognising this cobblestoned old town as the capital of the country at the intersection of Europe and Asia.
- “Savannah”, “Augusta”, “Louisville”, and/or “Milledgeville”, then your information is out of date (and unfortunately, so is AI’s sometimes).
- “I don’t know”, then perhaps you need to do a little more research.
The problem is all these answers are wrong. As far as I am concerned, the capital of “Georgia” is the letter “G”. And therein lies the problem.
As accountants, we are relying on artificial intelligence (AI) tools more and more. Love it or loathe it, AI is becoming ever more integrated into our work and personal lives. We need information faster than ever before. However, as well as the required results being timely, they need to be accurate, unambiguous, and fit for purpose.
As a profession, we need AI for a variety of tasks. These include:
- Automation: If you have ever experienced the unbridled joy of copying data from one place to another, removing excess spaces, deleting duplicates, checking for omissions, filtering out errors, and sorting alphanumerically in order to undertake data entry, prepare invoices, process payroll, or reconcile bank statements on a daily basis, then you will appreciate AI.
Whilst instructing AI to “prepare my daily report for my manager” might cause software consternation, a better worded, “Review all incoming invoices dated September 2025 from my MIS folder, extract vendor information and invoice amounts, and update the accounts payable ledger automatically” may prove much more fruitful. - Insights: Large language models (LLMs) are highly useful for identifying anomalies; extracting clusters, patterns, or trends; and conducting various other types of systematic analysis such as ratio analysis or breakeven calculations. “Analyse the following data” may provide hit-and-miss results, but “Review monthly expense data for the past three years and identify any months where travel expenses and/or personal reimbursements exceeded more than 20% of the employee’s average monthly salary” may deliver the information you need. Succinct is not always best.
- Communication: LLMs are great for matters of language. Effective communication is essential when trying to deliver a key message to stakeholders with varying levels of financial awareness. “Summarise our quarterly results” is far too vague, but opting instead for, “Draft a plain language summary of the Group Income Statement and Balance Sheet for the past quarter, highlighting the key changes and resulting recommended actions management should take (explaining why)” will give you a much better, more reasoned output.
- Learning: The worlds of tax, audit, assurance, accounting, and finance are constantly shifting. New accounting standards, tax legislation, and compliance requirements are rife. We need training. AI can be your personalised trainer. “Explain the new standard to me” is appallingly ambiguous, but “Summarise the main implications of the new international accounting standard IFRS 16 on leasing and explain how this will impact my mid-tier manufacturing firm in Wisconsin, if at all,” will get you the information you require. Let’s not even discuss the myriad ways AI can deliver such a summary.
- Effective visualisation: They say a picture is worth a thousand words. Using the right chart can make financials easier to understand: The story becomes more accessible and compelling with the right graph. No one wants trend analysis displayed on a pie chart, but “Create a dashboard with a pie chart of expense categories for the current quarter, a line graph of monthly cash flows over two years extrapolated for the next six months, and a heat map showing overdue receivables by client on a geotemporal basis” (thanks to years of Star Trek for enhancing my technobabble) may produce vivid, easy-to-follow, results.
- Fraud detection: LLMs can readily spot irregular patterns and proactively highlight areas of concern. Targeted, unsampled forensic analysis is now possible with AI. “Look for anomalies” may be what a character in a TV series might say to an unengaged subordinate, but the real-life professional may be more forthcoming: “Review all company credit card transactions for the past six months, highlighting all duplicate payments, split transactions just under approval limits, or any and all payments over $2,000 to new vendors not yet approved on the suppliers database.”
All of the above demonstrate the need for prompt engineering. Whether it is an art or a science is debatable, but this is when you can craft clear, articulate, unambiguous, and effective instructions, such that AI can deliver the results you require in a minimum number of iterations. This minimum number will not always be one: As a response is recorded, this may lead to follow-up or clarifying questions — iterations. An accountant must remain curious. A bad prompt may lead to vague or incorrect answers and potentially inappropriate decisions; a good one may simply lead to more questions.
AI is your assistant, but it is a very literal assistant. Therefore, I recommend you consider five key elements for a good prompt:
- Task: What exactly do you require? Explain why. Use unambiguous, plain language.
- Persona: Allow AI to understand the context, such as assigning a specific role (eg, “I am chief financial officer of a large multinational company, and I need to understand the following …”).
- Format: Make it clear how you want the results/response to be displayed. This may be in the form of a dashboard, a table, summarised bullet points, in Shakespearean iambic pentameter, or formal text.
- Example(s): Providing a sample of expected output(s) will help. Providing two to three illustrations may allow AI to home in on the expected output, but more may confuse, unless they are consistent and sufficiently detailed. Quality matters more than quantity.
- Constraint(s): Set restrictions or rules (eg, “top three factors”, “focus on the commercial aspects only”, “ensure data is up to date”).
Even with a good prompt, AI may not get it right the first time (you should always check facts to ensure current veracity). Prompt iteration is not an admission of failure. If the answer doesn’t feel right, consider these options:
- Ask the question in a different way.
- Break it down into smaller parts.
- Add more examples.
- Narrow the focus.
Essentially, good prompt + constructive feedback = efficient working.
For example, imagine you have been asked to put a training course together on AI. Do you think you will get a good result by using the prompt, “Write a one-day course outline on AI”? I think not.
It would be better to outline the context, the learning objectives, who the course is for, and what their roles are/how they will be using AI. Through a series of prompts, you could derive a course outline, one major topic at a time, with examples. However, it takes patience, skill, feedback, iteration — and determination.
You will develop your own style as you fine-tune your rapport with your AI tool of choice, but consider the overarching styles that are prevalent:
- Zero–shot prompting: Here, a question is asked with no examples, eg, “Summarise this report”. This is prone to inappropriate responses, misunderstandings, and so on.
- One–shot prompting: In this instance, just one example is provided to help AI understand what is expected. An example could be, “Here is a good example. Now write one like this.” This can work on occasion.
- Few–shot prompting: This requires two or more examples, providing greater control, detail, direction, and contrast. You are more likely to obtain reliable answers, and this is probably the optimum method for business tasks such as information extraction, report writing, and trend analysis.
Prompt engineering is the new coding for the AI world, and it is a critical skill for all accountants. It will take time to master, but the benefits will be immense, such as greater report quality, more timely analysis, better risk detection, and more accurate forecasting.
It will only improve with practice. So go practise.
Liam Bastick, FCMA, CGMA, FCA, is director of SumProduct, a global consultancy specialising in Excel training. He is also an Excel MVP (as appointed by Microsoft) and author of Introduction to Financial Modelling and Continuing Financial Modelling. Send ideas for future Excel-related articles to him at liam.bastick@sumproduct.com. To comment on this article or to suggest an idea for another article, contact Oliver Rowe at Oliver.Rowe@aicpa-cima.com.
LEARNING RESOURCE
Core Concepts of Artificial Intelligence for Accounting Professionals
Unlock the power of artificial intelligence with our comprehensive AI Core Concepts course. Gain a solid foundation in AI components.
COURSE
MEMBER RESOURCES
Articles
“Financial Modelling With AI: Part 3”, FM magazine, 11 June 2025
“Financial Modelling With AI: Part 2”, FM magazine, 2 June 2025
“Financial Modelling With AI: Part 1”, FM magazine, 23 May 2025
