Sorry I did not see the vision
An anonymous reader of your old writing (rumour says it’s Sabrina Carpenter)
Thank the Lord
The AI-assisted you has risen
If you write for regulators, boards or clients, AI can speed you up (and get you in trouble). This rubric shows how to use AI responsibly in your writing without risking your credibility or voice.
Context
You don’t have to take my word for it; we can look to the news for AI case studies. Recently, Deloitte was exposed for using undisclosed AI (without checking the output) in a $440,000 Government report. They ended up correcting the report and disclosing the AI, but it was a costly mistake.
Deloitte provides a specific case study on the direct cost of AI, but it is also vital to consider the environmental costs and reports of stolen IP. We will discuss how to approach AI with a critical eye, but it is also worth critically thinking about whether a task needs AI at all, given its issues.
For financial services professionals, these issues highlight the importance of disclosure, verification, and accountability when AI tools are used in preparing client advice, compliance reports, or Statements of Advice.
Doom and gloom, over; let’s get to the rubric.
Stage of Writing: Match the Tool to the Moment
At different stages of the writing process, AI can be used for different tasks.
| Stage | Possible AI Use | Calibration Notes |
| Ideation | Brainstorm topics, generate counterpoints | Stay in charge of what matters, and why |
| Outlining | Structure and flow suggestions | Get variants and make the final decision yourself |
| Drafting (Core) | Light boilerplate or low-stakes sections | Voice and insight must be yours |
| Drafting (Technical) | Summarise regulation | Cross-check everything |
| Editing | Clarify, condense, or improve transitions | Keep control over rhetoric and tone |
| Proofing | Grammar and polish | Accept if aligned with your style |
Cognitive Load
AI is better at supporting tasks with low- to medium-complexity. If the task demands judgment, interpretation, or legal nuance, AI’s role is strictly supportive.
Don’t let it write your argument or decide your message!
Fidelity Threshold: How High is the Cost of Error?
| Fidelity Level Required | AI Use Guidance |
| High (e.g., case law, legislation) | Use for reference, but verify rigorously! |
| Medium (e.g., interpreting reports) | Use with judgment; AI may misinterpret data |
| Low (e.g., tone smoothing) | Go for it, but review for consistency and tone |
Counterpoint Stress Testing
Use AI to break your own frame. The goal isn’t to have AI validate your view, but to work as a tool to push your thinking towards stronger arguments.
Prompt it to:
- Generate counterarguments
- Highlight implications
- Simulate opposing stakeholder positions
Ask questions like:
- “What would a regulator push back on here?”
- “What assumption am I not interrogating?”
- “How might this framing be misunderstood?”
Use this to improve the resilience and depth of your position.
Feedback Fidelity Score (FFS)
Rate AI suggestions across three axes:
| Axis | 1–2 | 3 | 4–5 |
| Relevance | Off-topic | Vaguely useful | Direct hit |
| Fidelity | Inaccurate | Sort of there | Factually grounded |
| Style Alignment | Robotic, weird vibe | Close but not there | In voice |
Only adopt edits scoring 4–5 across the board. AI is helpful, but don’t use it at the risk of your voice and credibility.
Final Filter: The Expert Room Test
Before incorporating any AI-generated suggestion, ask yourself:
“Could I defend this in front of a regulator, peer reviewer, or expert panel?”
If not, it doesn’t belong in your work.
Closing Thoughts
Used well, AI doesn’t write for you. It argues with you. It reveals assumptions, tests structure, and forces clarity. But only if you keep your hand on the wheel. Treat it as an intellectual sparring partner that helps you move faster by thinking harder.
This rubric isn’t a checklist to follow mindlessly; it’s a way of working that demands discernment, not dependence. Don’t expose yourself to unnecessary risk, but approach AI with a critical eye and belief in your own capacity.
If you’d like help navigating the complexity, reach out to Assured Support today.
If you enjoyed this article, you might also like:
- The Hidden Risks of AI Risk Frameworks
- The Green Cost of AI: Can AFSLs Balance It All?
- AI in Financial Advice: Efficiency Gains, Compliance Risks, and Cognitive Costs
- The Hidden Risks of AI: ASIC’s Review of Licensees’ Embrace of Artificial Intelligence
Tool Guide
It’s a case of trial and error to find an LLM that suits your needs, but here is a table of some tasks and tools to get you started.
Keep in mind that every tool will need you to check the output!
| Writing Task | Best Tool | Why It’s Best |
| Ideation / Brainstorming | GPT-4o (ChatGPT), Claude 2.1 | Strong creative thinking, analogy generation, and reframing complex problems across disciplines. |
| Outlining / Structuring Arguments | GPT-4o, Anthropic Claude 2.1 | Excellent at hierarchical logic and organising complex documents |
| Technical Summarization (e.g. ASIC reports) | Claude 2.1, Perplexity.ai Pro | Claude’s long context memory + Perplexity’s web access = accurate-ish summaries |
| Drafting Low-Stakes Content (Intros, bios) | GPT-4o (ChatGPT), Gemini 1.5 Pro | Fluent, varied style generation; good at tone-matching |
| Editing for Clarity & Flow | GPT-4o, Claude 2.1 | Strong at restructuring arguments, improving transitions, and refining tone without losing meaning |
| Tone/Diction Calibration | GPT-4o (with custom instructions) | Best at mimicking professional or institutional voice. Custom GPTs outperform default settings |
| Grammar & Proofing | Grammarly Premium + GPT-4o | Grammarly = fine-grained grammar; GPT-4o = stylistic finesse |
| Stress-Testing Arguments (Devil’s Advocate) | Claude 2.1, GPT-4o | Claude is tactful and precise in opposition; GPT-4o is better at broad scenario testing |
| Simulating Expert Pushback (e.g., Regulator) | GPT-4o, Claude 2.1 | Use prompt engineering: “Argue as if you are ASIC…” |
| Factual Cross-Checking / Source Validation | Perplexity.ai Pro, Bing Copilot | Retrieval-augmented tools. Useful for real-time data, citations, and triangulating regulatory positions (but validating is for humans) |
| Simplifying Legal/Technical Language | Claude 2.1, GPT-4o | Claude is better at “explain like I’m 12”; GPT-4o better at “smart but clear” |
Frequently Asked Questions
Because regulatory and board documents often contain legally binding advice or disclosures, AI errors or undisclosed use can compromise compliance, accuracy, and trust.
It refers to using AI-generated content without informing the client, regulator, or reader, which may breach transparency or professional conduct standards and undermine your expertise.
Assess both the complexity of the task and the cost of error. Use AI for low-risk drafting or editing, but retain human judgment for analysis, advice, and interpretation.
A simple pre-publication question: “Could I defend this AI-assisted section before a regulator or expert panel?” If not, it shouldn’t be included.
GPT-4o and Claude 2.1 perform well for drafting and tone calibration, while Perplexity.ai Pro helps with citations. Always verify outputs manually.
