Generative and deterministic AI fail in very different ways. Knowing which one you're using is becoming a professional skill.
AI is not one thing. For advisers, two types matter, and the difference is not well discussed by AI providers.
Generative AI writes and reasons in language. Deterministic AI computes against fixed rules. Both are useful. Neither is interchangeable. An adviser who can't tell them apart is one bad output away from a mistake they can't defend.
Two machines, one label
Generative AI includes ChatGPT, Copilot, Gemini and Claude. These large language models predict the most likely next word, based on patterns in their training data. Ask the same question twice and you can get two different answers, because they sample from probabilities rather than follow fixed rules. Training, grounding and context have lowered the error rate quickly, but the errors are still there.
Deterministic AI is older, quieter, and already embedded in your practice. Your projection engine is deterministic. So is your fee calculator, your Xplan modelling, your Centrelink calculator, and the decision tree your compliance team built five years ago. Same input, same output.
The practical distinction: generative AI estimates patterns in language and meaning. Deterministic AI computes using formulas and rules you can inspect. That difference explains why each tool succeeds and fails where it does.
Why this matters for your practice
The risk is not AI itself. It is casting the wrong type of AI for the job.
Drafting a file note from a meeting transcript? Generative AI is exceptional. Summarising a fact find, restructuring messy notes, drafting a client email — this is what generative AI does best.
Calculating a client's projected retirement balance? Comparing two product fee structures? Modelling a transition to retirement? Do not outsource that to a chatbot. A generative model can sometimes produce a plausible projection. You should not rely on one as your primary calculation engine in regulated advice, because nothing guarantees the answer ties back to the underlying formulas, assumptions or legislation.
Generative models hallucinate. They produce confident, plausible answers that are wrong. You cannot prompt your way out of this. The system is doing what it was built to do: generate fluent text that looks right, not verify every number against the Corporations Act or your SOA assumptions.
Asking a writing tool to do a maths job that must be explainable and traceable is dangerous in regulated advice. The output will look right and be wrong, which is the most dangerous failure mode there is.
The professional skill is not "using AI". It is matching the tool to the job. Generative AI for language work. Deterministic AI for calculations and rule-bound modelling.
Spot check: is this job generative or deterministic?
Use this before you reach for a tool.
Generative — client communication and records:
- Drafting a file note from meeting transcripts
- Drafting or rephrasing client emails and review letters
- Turning internal jargon into client-friendly explanations
Deterministic — advice maths and modelling:
- Calculating retirement projections and scenario comparisons
- Comparing product fees, costs and projected balances
- Running contribution, TTR and pension optimisation strategies
- Applying tax, Centrelink or product-specific rules
When in doubt: if the job is about words, reach for generative AI. If the job is about numbers or rules, reach for a deterministic engine, then use generative AI to explain the result in plain language.
What ASIC is watching
In October 2024, ASIC released REP 798, Beware the gap: Governance arrangements in the face of AI innovation. AI adoption is outpacing what the regulator can keep up with.
ASIC's core message is that the regulatory framework is technology neutral. It applies to AI the same way it applies to a calculator, a spreadsheet, or a paraplanner. A model being complex or branded as "AI" does not lower the bar.
Best Interests Duty does not care which type of AI produced the output. Your name is on the advice. You need to know which machine did the work, what role it played, and whether that machine was the right one for the question.
Two questions before you hit send
Before you send anything to a client:
- Did I use the right type of AI for this job?
- Can I explain how it got to the answer?
If you cannot confidently answer both, do not send it. Fix the process first.
The AI question for advisers is not if. It is which one, and for what.
Peter Worn is Joint Managing Director of Finura and Co-Founder of Advice Designer.
