Our method
The same five steps run through every article, learning module, skill file, recipe and tool on this site. They are not a theory of AI. They are what has to happen between an AI draft and a decision someone can stand behind in regulated Australian work.
This method was in use before it had a name. It is the shape of every workflow we publish, from a determination-drafting project space to a board paper, and it is the reason our headlines keep landing on the same line: AI can build the thing, it cannot make the call. Naming it means you can hold us to it, and reuse it, rather than rediscovering it article by article.
The steps run in order and each one earns its place. Skip the first and the rest are unsafe. Skip the last and you cannot prove any of it happened.
Nothing that identifies a person goes into a general-purpose AI tool. Names, claim and file numbers, dates of birth, addresses, employee identifiers and free-text clinical or investigation detail come out before the prompt goes in, replaced with placeholders you can map back afterwards. This is the step that makes every step after it possible: once the material is de-identified, the same tool that was too risky to touch a claim file becomes safe to think with. It is also the step people skip first when they are busy, which is why it is step one rather than a footnote.
A model recalling an Act, a prudential standard or a regulator guide from memory is not a source. Give it the current compilation, the guidance, your own policy and procedure, and the de-identified file, then instruct it to work only from what you supplied and to say plainly when the answer is not in there. Grounding is what turns a plausible answer into a traceable one. It also changes what a wrong answer looks like: an ungrounded model invents a section number, while a grounded one tells you the document does not cover the point.
The model assembles the picture. It does not make the call. Liability, notifiability, reasonableness, eligibility, the fairness of an outcome and anything else the law or your own delegations reserve for a person stay with the person who holds that authority, and they exercise it on the evidence rather than on the draft in front of them. The practical test is simple: if a regulator or a tribunal would ask who decided, the answer has to be a name, not a tool.
Every load-bearing legal, statistical or product claim is checked against the authoritative text: the registered compilation of an Act, the published paper, the regulator page, the model card. Paraphrase that has drifted from the source gets pulled back to it, and a claim that cannot be traced does not run in a weaker form, it comes out. This is the step that catches the confident, well-formatted, entirely wrong sentence, which is the failure mode that matters most in regulated work.
Record what went in, which tool produced the draft, what a person changed, and who signed it off. A log is not bureaucracy for its own sake: it is the difference between saying you kept a human in the loop and being able to show it, months later, to someone entitled to ask. It also makes the method improvable, because a trail of what the model got wrong is the only reliable guide to where the controls need tightening.
One published worked example per profession. Each links to the article it comes from, so you can read the full workflow, the prompts and the governance line rather than take our word for the summary.
Workers compensation
A section 57 examination requirement became a reviewable determination on 14 June 2024 and must comply with the Comcare guide for arranging assessments and examinations. AI assembles the referral, the chronology and the decision record from a de-identified file. The decision to require the examination stays with the delegate, and the record exists because that determination is reviewable.
Section 57 Examinations: AI Can Build the Referral, Not Make the CallWork health and safety
Three uses hold up: de-identified trend and leading-indicator analysis, causal hypotheses for a single incident, and drafting investigation write-ups with the gaps left blank. One never does: auto-classifying whether an event is notifiable under sections 35 to 39. That judgement belongs to a competent person, and the de-identification rule is what makes the other three safe.
AI for Incident Analysis and Leading Indicators: A Human-in-the-Loop WHS PlaybookGovernance, risk and compliance
AI can triage complaints, summarise files, draft acknowledgements and surface possible systemic issues. Under ASIC RG 271 the enforceable core stays human: the 30 calendar day clock, the reasons an internal dispute resolution response must give, the systemic issue call and the fairness of the outcome. The article works through a drafted rejection that did not survive review, which is the human step doing its job.
AI in Complaints Handling: What RG 271 Reserves for a PersonHuman resources
AI structures the request, identifies the missing facts, builds the consultation agenda, compares alternatives and drafts the written response from verified inputs. It does not decide eligibility, whether business grounds are reasonable under section 65A, or whether a refusal is lawful. HR and the authorised manager consult, weigh the evidence and own the decision.
AI Can Map a Flexible Work Request. It Cannot Decide ReasonablenessLeadership
Let AI prepare, observe and compress the meeting, and never let it certify. A named person with authority confirms what was decided across a six-field record covering decision, evidence, conditions, authority and ownership, dissent, and execution. AI drafts those fields from approved notes and flags anything missing as not recorded, for the chair to resolve before the record is issued.
The AI Joined the Meeting. The Decision Record Still Belongs to You.The skill file library ships the method as loadable instructions, the workflow recipes run it as fixed procedures for recurring tasks, and the platform configurations set it up inside ChatGPT, Claude, Microsoft 365 Copilot and Gemini Notebook. The same discipline governs how this site is written: see the worked corrections in our editorial standards.
General information and education only. Not legal, compliance, financial, or professional advice. The method is a working discipline, not a compliance certification: your organisation's policies, your delegations and the relevant regulator guidance govern any workplace use.