Smart AI for Smarter Accountants: Use AI Responsibly and Intelligently
This article will count 0.25 units (15 minutes) of unverifiable CPD. Remember to log these units under your membership profile.
The Chartered Institute of Business Accountants (CIBA) recently submitted a detailed proposal to the South African Revenue Service (SARS), urging the agency to invest in artificial intelligence (AI) tools, subsidised platforms, and skills development for tax practitioners to counter growing technological asymmetry between SARS and the advisory profession. SARS recently confirmed it will spend R3.5 billion over the next three years on modernising its systems, deploying artificial intelligence, and automating compliance enforcement.
AI offers immense power to transform your daily tasks as a business accountant. Used correctly, it can significantly boost your productivity and effectiveness. It can help you draft documents faster. It can also assist with complex research. However, AI also carries risks. The primary risk is unreliability or "hallucination"—AI making up facts. You are ultimately responsible for all advice delivered with or without the use of AI. Hallucination by AIs or inadequate supervision of AI’s work could lead to incorrect client advice. You can avoid these risks by adopting the simple techniques detailed below. This guide will show you how to use AI responsibly and smartly.
Practical Techniques for CIBA Accountants
Here are the techniques you can apply immediately, ensuring high quality and reliable outputs from your AI tools.
Selecting the Right AI for the Job
Use the AI, or combination of Ais, most suited to your purpose. Different AIs excel at different tasks (Large Language Models – LLMs – for writing, Researchers for verification, Creators for visuals).
➡️ How-to do it:
If you need to research a complex section of the Income Tax Act, use a research-focused AI like Perplexity first to get verified data and sources. Then, use an LLM (like ChatGPT) to draft the client advice using that verified information.
The CRISPE Model for Clarity
Use the CRISPE model to ensure your prompt is complete: Context, Role, Intent, Scope, Parameters, Example. This structure gives the AI everything it needs to perform a task correctly.
➡️ How-to do it:
You need the AI to create a cash flow forecast summary for a business loan application. Instead of "Draft a forecast," use all six CRISPE elements:
Context: "The client is a small, profitable retail store in South Africa. They are applying for a R500,000 short-term loan for inventory expansion."
Role: "Act as a CIBA Business Finance Analyst reviewing the loan application documents."
Intent: "Draft a three-page executive summary focusing on the client’s projected cash flow for the next 12 months to support the loan application."
Scope: "The forecast must show monthly cash inflows (sales) and cash outflows (rent, salaries, inventory). It must clearly state the minimum cash balance projected in the next year."
Parameters: "The summary must be written in a persuasive but realistic tone, suitable for a commercial bank's credit manager. All currency figures must be in South African Rand (ZAR)."
💡 Example: Reference this sentence structure for the concluding statement: “Based on conservative assumptions, the projected cash flow shows that the business will maintain a minimum monthly cash surplus of R[insert amount], demonstrating ample capacity to service the R500,000 loan.”
Prompt Sequencing
Break down complex accounting tasks into smaller, manageable steps, and sequence the tasks for the AI. This directs the AI through a multi-stage process.
➡️ How-to do it: For an impairment calculation:
✏️Prompt 1: "Calculate the recoverable amount (higher of value-in-use and fair value less costs to sell) for the following asset..."
✏️Prompt 2: "Based on the recoverable amount from the previous step, determine and state the journal entry required under IFRS to record the impairment loss."
Chain of Thought
The Chain of Thought technique is an instruction that forces the AI to show its logical working before providing the final answer. This makes the output auditable.
➡️ How-to Application: For a complex deferred tax calculation:
✏️Prompt: "Use the Chain of Thought technique to first explain the steps for calculating deferred tax on property, plant, and equipment (PPE) under IFRS. Then, calculate the deferred tax liability/asset using the figures I provide."
Prioritise Your Instructions
LLMs typically put more emphasis on more recent information. Therefore, put your most important instructions at the end of the prompt.
➡️ How-to Application:
✏️ Prompt: "Draft a response to a client's query about IFRS 16 lease classification. Keep it under 200 words. Include a warning that final accounting decisions must be approved by the client's registered auditor." (the warning, being the most important instruction is at the end of the prompt.)
Giving "I Don't Know" Permission
Explicitly give the AI permission to say that it does not know. This is a crucial safety measure to prevent the AI from "hallucinating" (making up facts) when it encounters a gap in its knowledge.
➡️ How-to Application:
✏️ Add this instruction to your prompt: "If you are unsure of any part of the answer, or if the SARS position is unclear, you must state 'I cannot verify this information' rather than speculating."
Meta Prompting
Use the AI itself to suggest better prompts for you. This is AI-assisted quality control for your inputs.
➡️ How-to Application:
✏️ Prompt: "I want to use you to review the wording of an accounting policy note for revenue recognition. What is the best prompt I can use to ensure you check for compliance with IFRS 15 and that the language is unambiguous?"
Source Verification
Always ask for the sources. If your AI cannot provide sources, then review the output with an AI that provides sources (e.g., Perplexity). Absence of a source is a red flag on reliability.
➡️ How-to Application:
✏️ After an LLM generates a summary of the latest South African budget changes, ask: "Provide the specific URL or official SARS document reference for each claim you made above."
Confidence and Expert Self-Review
Ask the AI to assess its own output. Ask AI to state its confidence or certainty level with the output and/or ask for an expert review by AI itself.
➡️ How-to Application:
✏️ After the AI drafts a client email on provisional tax, ask: "Rate your certainty on the accuracy of the tax deadlines provided (1 to 5). Then, acting as a Senior Provisional Tax Expert, critique this email and with particular focus on the areas where your confidence is lower than 95%"
Iterate
Iterate means to refine your prompt and run the task again until the result is perfect. Don't accept the first draft the AI gives you; refine it until it meets your professional standard.
➡️ How-to Application: You asked the AI to draft a letter about an asset disposal, but the tone was too casual.
✏️ First Prompt: Draft a client letter about asset disposal. You find the result to be too casual.
✏️ Reiterated Prompt: Draft the client letter again. Use the prior output, but change the tone to strictly formal and professional, suitable for a regulated CIBA accountant.
Cross-Check (or Critique)
Use other AIs to cross-check or critique the initial output. This leverages the strengths of different models to find errors or improve clarity.
➡️ How-to Application:
✏️ Use your primary LLM (e.g., ChatGPT) to draft an executive summary of the annual financial statements. Then, paste the summary into a second LLM (e.g., Gemini) with the prompt: "Critique this executive summary for clarity, consistency of financial logic, and professional tone, specifically from the perspective of a CIBA-qualified analyst."
Inject Your Authentic Voice
AI output can sound generic. Modify the output to inject your authentic voice. This is essential for maintaining your relationship with your clients.
➡️ How-to Application:
If the AI uses overly formal or complex language (e.g., "commence an exhaustive examination"), rephrase it to your familiar, simpler language (e.g., "start a detailed review").
Human Expert Review
All outputs should pass through human compliance review and be supported by documentary evidence, consistent with CIBA’s quality assurance protocols for trusted advisors. AI is a powerful assistant, but it is not a replacement for professional judgment. Always ensure expert reviews by human experts (yourself and senior colleagues) of critical outputs before signing off.
➡️ How-to Application:
You use AI to draft an annual financial statement disclosure note. You must review and sign off on the final output yourself to ensure it aligns with the client's specific circumstances and your CIBA professional obligations.
Avoid Pleasantries
Avoid using polite phrases like "Please" or "Thank you" in your prompts. Ais are emotion neutral. Adding pleasantries just adds more word for the AI to process, reducing efficiency.
Conclusion
AI is here to stay, and it represents a significant opportunity for every CIBA accountant to work smarter, not just harder. By diligently applying these techniques—from using the right tool to auditing the final output—you move beyond simple AI usage. You will ensure your work remains accurate, reliable, and compliant, all while boosting your productivity and maintaining your professional integrity. Embrace AI intelligently, and watch your practice thrive.
