Teaching Audit & Accounting Teams to Use ChatGPT and Copilot Safely

This industry requires a higher standard of data hygiene than almost any other. The focus here is on leveraging AI for efficiency without compromising client confidentiality or falling for "math hallucinations."

How to start, what to type, and how to get value fast—while maintaining professional skepticism and data secrecy.

1) Start Here: The 3-Step Method (Hide → Ask → Check)

Financial data is the holy grail for hackers. Treat AI like an intern who is brilliant at writing but terrible at keeping secrets and bad at mental math.

Hide Client/Entity data (Anonymise)
Ask for a specific structural or textual output
Check for "Math Hallucinations" and Regulatory Accuracy
- Never ask the AI to calculate the sum of a ledger.
- Replace specific figures, client names, and tax IDs with tokens:
  - Company ABC, $5,000,000, Account 1234 → [CLIENT], [MATERIALITY_THRESHOLD], [BANK_ACC]

One Rule: Zero Trust on calculation. AI is a language model, not a calculator.

2) What to Type: Prompt Anatomy (copy this pattern)

Role/Task + Context + Constraints + Format

You are a Senior Audit Manager.
Task: Draft a management letter comment regarding a control deficiency.
Context: During our walkthrough of [PROCESS], we noted that the [JOB TITLE] can both create and approve vendor master data changes. This is a segregation of duties conflict.
Constraints: Professional, constructive tone. Cite the risk of fraud. Do not reference specific employee names.
Format: Observation, Risk, Recommendation.

3) First 10 Prompts to Try (beginner-safe)

1. The Excel Formula Fixer (Most Popular)

"I need an Excel formula to look up values in Column A against Sheet 2, but only if Column B contains 'Active'. If not found, return 'Check'. Explain the syntax."

2. Accounting Standard Summariser

"Summarise the key criteria for lease recognition under IFRS 16 (or ASC 842) for a non-specialist. Highlight the exemptions for low-value assets."

(Note: Always verify against your firm's technical guidance).

3. Client Request List (PBC) Drafter

"Draft a polite but firm email to a client controller attaching the outstanding 'Provided by Client' (PBC) list. We need these items by Friday to meet the filing deadline. Context: We have asked twice already."

4. Control Narrative Generation

"Turn these rough interview notes into a formal process narrative for the 'Procure-to-Pay' cycle.

Notes: [REDACTED NOTES - e.g., 'PO created by dept head, sent to finance, matched with invoice...']"

5. Risk Assessment Brainstorming

"Act as an external auditor. What are the common fraud risks associated with revenue recognition in a [SPECIFIC INDUSTRY, e.g., Construction] company using percentage-of-completion?"

6. Policy Comparison

"Compare the text of this [DRAFT T&E POLICY] against standard best practices for a mid-sized tech company. Identify 3 missing controls regarding approval limits."

7. Drafting Review Notes

"Rewrite these rough review notes to be more coaching-oriented and less aggressive for a first-year associate.

Draft: 'This is wrong. Fix the tickmarks. You missed the tie-out.'"

8. Macro/VBA Writing

"Write a VBA macro for Excel that formats the selected range: bold headers, freeze top row, and highlights any cell containing the word 'Error' in red."

9. Flux Analysis Explanations

"I have a variance of +20% in Legal Expenses. Write a concise explanation for the board report based on these facts: [REDACTED FACTS - e.g., settled two IP lawsuits, one-off merger costs]."

10. "Explain it to a Client"

"Explain the concept of 'Deferred Tax Assets' to a client CEO who has no accounting background. Use a simple analogy."

4) Live Workflow (The Audit Cycle)

Planning:

Use Prompt #5 to brainstorm risks before the planning meeting.
Use AI to draft the agenda for the client kickoff meeting.

Fieldwork:

Excel helper: Keep a chat window open to debug formulas or write macros to clean messy GL dumps (Prompt #1 & #8).
Documentation: Paste redacted notes from walkthroughs and ask AI to "Format this into a clear memo" (Prompt #4).

Reporting:

Use Prompt #2 to refresh your memory on disclosure requirements (then verify!).
Draft the Executive Summary of the Audit Findings report.

5) Before/After (see the difference)

❌ Unsafe

"Analyze this GL dump for Company X. Why is account 4000 showing $4.2M?"

(Risk: Uploading financial data, relying on AI for math).

✅ Safe (after Hide → Ask → Check)

"I am analyzing a GL for a retail client. Revenue has increased 15% but Cost of Sales only 2%.

List 5 potential accounting anomalies or operational reasons that could explain this margin expansion to investigate further."

6) Common Mistakes (and quick fixes)

Using AI for Math: Never ask "What is 10% of this list?" It often hallucinates numbers. Use Excel for math, AI for text.
Copy-Pasting Client Emails: Even internal emails contain confidentiality. Summarise the issue, don't paste the thread.
Blind Trust on Tax Law: Tax codes change yearly. AI training data has a cutoff date. Always cross-reference the tax year.
Lazy Review: AI sounds confident even when it's wrong about a standard (e.g., mixing up IFRS and US GAAP).

7) What AI Is Great At vs. Not Great At

Great at:

Excel Syntax & VBA: It is better at Excel than most humans.
Drafting Text: Memos, emails, policy descriptions.
Synthesizing Concepts: Explaining complex standards simply.
Formatting: converting messy bullet points into tables.

Not great at:

Calculation: Do not use it to add, subtract, or audit figures.
Recent Regulations: It may not know a law passed last month.
Judgment: It cannot determine if a misstatement is "material" – that is your judgment.

8) Mini-Checklist (stick near your screen)

[ ] I removed Client Name and Entity ID.
[ ] I removed specific dollar amounts (or replaced with placeholders).
[ ] I am using AI to generate words/code, not calculations.
[ ] I verified the accounting standard (IFRS/GAAP/Tax) cited by the AI.
[ ] I read the output for "hallucinated" facts before pasting it into workpapers.

9) Tools & Setup

Microsoft Copilot (Enterprise): Best for this sector as it integrates with the Office suite (Excel/Word) and usually offers better data protection guarantees for enterprise.
Internal Knowledge Base: If your firm has an internal AI trained on your own methodology manuals, use that for technical queries over public ChatGPT.

10) 15-Minute Team Starter

5 min: Review the "Math Hallucination" warning. Show an example where AI gets a calculation wrong (or explain why it happens).
7 min: The Excel Challenge. Everyone opens a messy spreadsheet. Use AI to write a formula or macro to clean it up (e.g., "Write a formula to extract the domain name from these email addresses").
3 min: Share the most useful Excel formula prompts discovered.