PREFACE

There was a recent BBC Travel article ‘’The perils of letting AI plan your next trip’’  about tourists misled into visiting places that don’t exist. The article demonstrates how easily AI can sound convincing while being wrong.

Executive Summary & Methodology

This guide is a reference guide for business users, on the use and dangers of AI tools. All content is grounded strictly in user experience, cross-model checking, and real-world test cases. It was drafted, for efficiency sake, with the help of AI but the tips / advice listed are based on my personal experiences from heavy ai usage as a hobby (cross checked and validated with multiple ai models)

1. Prompt Specificity and AI’s Bias to Consensus

AI will usually answer with whatever information you give it—without asking for clarifications unless prompted. However, even if your prompt is totally clear and specific, the AI may still ignore your instruction and provide the established or consensus answer instead of following your actual facts or request.

Example — Consensus/Establishment Bias:
I gave an AI model a passage from Genesis in ancient Greek and asked it to translate it into English. My prompt was clear and specific: “Translate this to English.” Instead of giving a direct translation, the AI repeated the wording found in established English translations, even though those did not match the original Greek. When I pointed out the error, it admitted it had relied on what previous translators had done, not on a literal translation from the text I provided.

This demonstrates that even a crystal-clear prompt may be overridden by the AI’s tendency to default to consensus, established science, or widely accepted answers and not strictly the data or instructions you provided.

Prompt Safety Checklist:

• Always instruct: “Feel free to ask clarifying questions before answering.”
• “No filling in of gaps, no generalisations, no guesses.”
• “I want a serious (e.g. financial) professional answer as an independent expert in the field.”
• “Give me your sources for your answers.”

2. Hallucination and Checking AI Sources

AI can hallucinate—meaning it may invent data, cite sources (and even tourist destinations!) that don’t exist, or make up details. Although hallucination rates have dropped, you must always check any sources provided.
When AI gives you a source make sure to check it out either by yourself or by inputting it in another model and asking for verification.

3. Misapplied Rules, Science, or Regulations

AI tries to be efficient, but that often leads to incorrect application of the sources or rules it finds. Ensure the AI has all the specific data needed. Provide precise facts, jurisdiction, and context for your scenario.

4. Shortcuts and Handling Long Lists or Big Data

AI is built to be efficient. When given long data, big lists, or large texts, most AIs—including ChatGPT—will often estimate the answer rather than process every detail. They may skip lines, claim to have checked all data, or generalise—when in reality they have only processed a subset. This is particularly dangerous for financial managers and accountants.

To overcome this:

• Always give data in small chunks.
• Ask the model to reproduce or echo the data it actually used.
• Request a step-by-step breakdown and proof for any answer based on lists or tables.

Note: Some models (like Perplexity and DeepSeek) are much better at this level of detail, but DeepSeek’s servers are often busy.

5. Model Comparison Cheat Sheet

6. AI Dependency & Double-Check Workflow

AI is powerful, but risky to depend on alone. If you need rely on it for critical advice (like tax or IFRS treatment), always cross-check the answers.

Method:

• Paste your prompt and the AI’s answer into a different model and ask: “Do you agree? Why or why not?”
• If there’s disagreement, paste that answer back into the first model.
• Repeat until both models converge on an answer.

7. Valid Use Cases for Financial Managers

(*Never upload sensitive / confidential data in AI)

8. Special Section: Google NotebookLM for Finance

NotebookLM (Gemini) is fundamentally different: it does not use external data, only what you upload. You can upload financial statements, reports, SOPs, or any internal document set, and NotebookLM will read, summarise, and answer questions only from those sources.
Example: Upload a set of consolidated financial statements. NotebookLM can walk you through the notes, explain changes, or compare with other files—without ever leaking data outside your private session.

9. Confidentiality & Professional Judgement

Never use AI to get information or advice about something you cannot later check for yourself. You must always be able to judge if the answer makes sense. Even after double-checking models, if your prompt isn’t precise enough, there is always a risk of error.

Privacy and Security:
ChatGPT-5, ChatGPT-4o, and other OpenAI business models (Team, Enterprise, Edu) do not use your data for training by default, and offer admin/user retention controls. Free/public versions may use your data for training unless opted out—never paste sensitive information unless you’re on a business plan. Perplexity Pro/Enterprise offers retention controls. Bloomberg Terminal AI features are only accessible inside the Bloomberg ecosystem and never train on your data.

10. Memory, Session Limits & Saving Insights

AI models do not actually read your whole chat—they operate from a limited memory buffer. For long or detailed projects:

• Break tasks into short, specific sessions
• Explicitly instruct the AI to save or summarise key insights
• Always export or save critical outputs for your audit trail
• When reaching capacity, start a new session with a summary of the previous chat in the beginning so that it can continue from where you left off. One chat windows does not have access to content in other chat windows. It only accesses what you instructed it to save in memory (for models that have this function).

Final Remarks & Call to Action

AI is a powerful support tool, but not a replacement for judgement. The key is not blind trust, but disciplined, informed use. Establish and regularly review internal guidelines for AI use. Document and export all critical steps for your own protection. Cross-check every major answer, and make a habit of asking for clarification and proof—not just for compliance, but for the integrity of your work.

The information in this guide is valid as of the date it was written. Parts may become outdated as AI evolves fast. Treat this as a living reference and review it periodically.

END OF DOCUMENT