Artificial intelligence can read a data room while you are boiling the kettle, but it cannot tell you whether a RTO is worth buying. A practical, evidence-based guide for buyers on what AI does well, where it fails and how to use it without fooling yourself or it fooling you.

As a special BONUS, at the end of the article there are prompts to use – and not use.

 

Why this matters now

Two forces are colliding in the Australian vocational education and training (VET) market. Buyers who understand only one will make expensive mistakes. The first is the most demanding regulatory setting the sector has operated under in a decade. The second is a technology that promises to compress weeks of due diligence into days. Buyers who understand both will make better decisions.

ASQA and state bodies are applying more scrutiny, not less. The 2025 Standards for RTOs took effect on 1 July 2025, and ASQA’s 2025–26 risk priorities centre on the integrity of qualifications and competency outcomes and on assurance in complex and international delivery models.^[1] ASQA rejected 27 renewal applications in the first half of 2025–26, more than in comparable earlier periods, and has reported that withdrawals have risen as some providers choose to exit rather than renew.^[2]

Why the big-deal numbers don’t transfer

AI is already part of how deals get done. Bain’s 2026 M&A Report found 45% of surveyed executives used AI tools in M&A in 2025, more than double the year before, and about a third now use it systematically or are rebuilding their processes around it.^[6] Deloitte found 86% of 1,000 US corporate and private equity leaders had adopted generative AI in dealmaking. They used it most for strategy (40%), target screening (35%) and due diligence (35%).^[7] McKinsey reports that generative AI has cut deal timelines by 10% to 30% and costs by roughly 20%.^[8]

Read those numbers carefully before you take comfort from them. They come from large, mostly overseas transactions, run by teams of lawyers, accountants and analysts whose job is to check what the AI produces. Your deal will be different. It’s usually a share sale between private parties, often under $2 million, and you’re probably doing much of the reading yourself. You can get the time savings. You usually won’t have the safety net that sits behind them.

What does transfer is this: AI is now on both sides of the table. Assume the information memorandum, the vendor’s answers to your questions and even your own adviser’s first draft may have passed through a chatbot. A polished document is no longer evidence that it’s accurate.

 

The evidence on what AI gets right and wrong

Any discussion of AI in a high-stakes purchase should start with what the research actually shows, rather than with vendor claims or anecdote. Four findings matter most for RTO buyers.

 

1. AI makes things up, and it sounds most certain when it’s wrong

Ask a chatbot for an RTO’s scope and it may give you a qualification the provider has never held, neatly formatted and stated with complete confidence. It may even be listed as “current” although the date of currency may have passed. That isn’t a glitch. It’s a known, measured behaviour.

A peer-reviewed Stanford and Yale study tested legal research tools sold specifically on the promise that they don’t invent answers. LexisNexis’s tool produced false or misleading answers 17% of the time. Thomson Reuters’ Westlaw tool did so 33% of the time. GPT-5, the general-purpose model most like what buyers actually use, got it wrong in more than two answers in five (43%).^[9] The tools are improving but as more information is available, the tentacles of the chatbot are wider and bring in spurious data. No data has been published as independent evidence that the problem has gone away.

This isn’t theoretical. In October 2025, Deloitte, one of the Big Four, agreed to partially refund the federal government for a $440,000 report that cited non-existent sources and contained a fabricated quote from a Federal Court judgment.^[10] If a Big Four firm can submit invented material to a federal department, a buyer working alone at 11pm can certainly put it into a purchase decision.

The pattern behind the failures matters most. A UNSW review found 87 Australian court and tribunal cases involving generative AI misuse. More than three-quarters of them involved people representing themselves, with no lawyer checking their work. In an RTO acquisition, that’s the buyer doing their own due diligence.

In a share sale, a fabricated fact doesn’t stay in a chat window. It goes into your price, your warranties and your loan application.

 

2. AI tells you what you want to hear

This is the most dangerous finding in this article.

A study published in Science in March 2026 tested 11 leading AI models. On average, they backed the user’s position 49% more often than people did. Read that again. Nearly 50% more often the AI will agree with you. Sounds like an echo chamber to me!

Even when users described harmful or deceptive conduct, the models endorsed it close to half the time. Worse, participants couldn’t tell the flattering answers from the objective ones, and they trusted and returned to the models that flattered them. This means, buyers are “trusting” AI more than advice or their own intuition and that is dangerous.

Picture a buyer three weeks into a deal, emotionally committed and keen to move, typing: “Why is this a good acquisition?” They will get a persuasive, well-structured answer. It reflects the question, not the business. If the question was, “Is this a good purchase? You are expecting AI to read your thoughts and emotions.

AI won’t be the voice that tells you to walk away. Make sure someone is.

 

3. Ask the same question twice and get two different answers

If a valuer valued the same business twice and gave you two different numbers, you’d never use them again. AI does this routinely.

Consumer chatbots are designed to vary their answers. Even when developers switch that variation off, the results still shift. In September 2025, research from Thinking Machines Lab showed the cause: commercial systems process your request alongside other users’ requests, and that subtly changes the calculation.^[14] It can be engineered out, at a cost in speed. The chatbots most buyers use make no such guarantee.

So a “yes” on Monday can quietly become a “maybe” on Tuesday, and you’ll only know if you asked twice.

A single AI answer is a sample, not a measurement. If your process doesn’t force consistency, nothing will.

 

4. AI makes you faster. Whether it makes you right is up to you.

The upside is real. The Productivity Commission considers multifactor productivity gains above 2.3% likely over the next decade, though it stresses considerable uncertainty.^[15] PwC found revenue per employee growing about three times faster in the industries most exposed to AI than in the least exposed.^[16]

Look at who is collecting those gains: large firms in finance and software, with teams of analysts, lawyers and auditors checking every output. PwC’s own economists say the growth has been driven by very large businesses. Australian SMEs are further behind, with 43% using AI at all.^[17]

The difference isn’t the tool. It’s the checking. Take away the checking and you keep the speed but lose the safety. Trust but verify.

Speed without verification doesn’t save you time. It gets you to the wrong answer sooner.

 

The bottom line

AI is a strong accelerator of reading, sorting, summarising and drafting. It is a weak and sometimes misleading substitute for verification, judgement and independent challenge.

Use it to read faster. Never use it to decide. The rest of this article shows how.

 

Pros and cons at a glance

Using AI across the buying process

The most useful way to think about AI in an acquisition is stage by stage. At each stage the question is the same: what can the tool do faster, and what must still be verified by a person against a primary source? The examples below are illustrative and assume the buyer uses a business-grade AI tool with appropriate data protections (see the section on what not to do).

A word on tools. The general-purpose assistants most buyers will meet include OpenAI’s ChatGPT, Anthropic’s Claude, Google’s Gemini and Microsoft Copilot, and many virtual data room and document review platforms now embed similar models. No product is immune to fabrication or variability.

 

Stage 1: Defining the brief and screening opportunities

Before looking at a single listing, a buyer should be able to state what they are buying and why. AI is genuinely useful here as a thinking partner, provided it is asked to challenge rather than confirm. A buyer might describe their capital, experience, preferred training packages, appetite for government funding or CRICOS exposure and intended level of involvement, then ask the model to identify the trade-offs and contradictions in that brief.

Screening is where consistency matters most. Rather than asking a chatbot whether each listing ‘looks good’, a disciplined buyer applies the same scoring template to every opportunity: registration tenure, scope composition, funding dependency, revenue concentration, international exposure, key-person reliance and regulatory history. AI can apply that template quickly to information memoranda and public data. The template, not the AI, determines what counts.

Stage 2: Public verification

Every RTO has a public footprint. www.training.gov.au records registration status, scope and history; ASQA publishes regulatory decisions; CRICOS-registered providers appear on the Commonwealth register; and the Department of Employment and Workplace Relations publishes international enrolment allocations for VET providers.

AI can help organise this research, but it should never be the source of it. Models frequently produce plausible but incorrect RTO codes, scope items, registration dates and decision histories. The deterministic method is to extract the data directly from the register, paste it in and ask the AI to compare it with the vendor’s claims. That turns AI from an unreliable researcher into a reliable comparison engine.

 

Stage 3: Reviewing the information memorandum and data room

This is where AI saves the most time. A typical RTO data room contains financial statements, management accounts, enrolment and completion data, Training and Assessment Strategies (TAS) for each qualification, assessment tools, trainer matrices, third-party agreements, funding contracts, policies and audit reports. Reading every document in full can take weeks.

Used well, AI can map documents against a checklist, identify what is missing, compare the TAS against scope and delivery mode, check that trainer qualifications align with the units delivered and flag inconsistencies between narrative and data. Bain describes acquirers using AI to assess a target’s cost base from public sources before a formal bid rather than accepting a seller’s assertions at face value.^[3] The RTO equivalent is cross-checking enrolment claims against completion data, funding statements and trainer capacity.

Stage 4: Financial analysis and normalisation

RTO financials have features that general-purpose AI handles poorly unless instructed. Fees received in advance create unearned income and a future training liability. Government funding is often paid on milestones. Owner remuneration and related-party expenses frequently need normalising. A share sale also brings historical liabilities.

AI can rebuild a profit and loss into a standard format, calculate ratios and trends, draft a list of proposed normalisation adjustments with reasons and stress-test the effect of losing a major funding contract or a large client. What it cannot do is decide which adjustments are legitimate. That remains the job of the buyer’s accountant and valuer, working from source records rather than summaries.

Stage 5: Compliance and quality due diligence

For most RTO buyers, compliance risk is the largest single variable in value. I do not agree that it should be and this is more a historical legacy thinking than current truths. AI can review policies and procedures against the structure of the 2025 Standards, check assessment tools for mapping gaps and compare stated course durations with the volume of learning. It can also help a buyer test one of the most current risks: how the target manages AI use by its own students.

This risk is now explicit. ASQA has recently issued Principles for the Responsible Use of AI in VET, which it says do not create new requirements but help providers meet existing obligations under the 2025 Standards; providers are encouraged to review their governance, privacy, workforce capability and assessment practices as the technology evolves.^[4]

 

Stage 6: Valuation, structure and negotiation

AI should not produce the valuation. It can, however, make a buyer a better consumer of valuation work: explaining methods such as capitalisation of future maintainable earnings, building sensitivity tables, testing how value changes with different earnings or risk assumptions and preparing questions for the valuer.

In negotiation, AI is useful for preparation and for drafting, particularly for mapping each identified risk to a commercial response. Buyers should be cautious about using it to ‘win’ a negotiation. A model optimised to please its user will happily produce aggressive positions that damage trust with a vendor whose cooperation is needed through transition.

 

Stage 7: Completion and the first 100 days

After settlement, AI is well suited to drafting updated policies, staff communications, a compliance calendar and a 100-day integration plan, and to summarising regulatory obligations triggered by the change of ownership. Here too the rule holds: drafts MUST BE reviewed against the actual Standards, funding contracts and ASQA guidance before use.

 

Creating unbiased, deterministic prompts

Because AI tends to mirror its user and varies between runs, the quality of the prompt has an outsized effect on the quality of the decision. ‘Unbiased’ means the prompt does not signal the answer the buyer wants. ‘Deterministic’ means the same inputs, processed the same way, produce results that are consistent enough to rely on, and any inconsistency is detected rather than ignored.

 

Ten rules for neutral, repeatable prompts in buying a RTO including prompts to use and not use.

1. Ask open, symmetrical questions. Replace ‘why is this a good buy?’ with ‘set out the strongest evidence for and against acquiring this business, with equal depth on each side’.
2. Assign a role with a duty of scepticism. For example: ‘Act as an independent due diligence reviewer whose job is to find what is missing or unsupported.’
3. Define the evidence standard. Instruct the model to use only the documents supplied and to cite the document and section for every factual statement.
4. Make ‘unknown’ an acceptable answer. Explicitly tell the model to write ‘Not evidenced in the material provided’ rather than infer, estimate or fill gaps.
5. Separate facts, assumptions and opinions. Require three labelled sections so assumptions cannot masquerade as findings.
6. Fix the output format. Specify tables, fields and scoring bands. Structured output is easier to compare across runs and across targets.
7. Use a locked template. Save the prompt, give it a version number and change it deliberately rather than improvising each time.
8. Run it more than once. Run important prompts at least twice (and in a second tool if available) and treat any material difference as a flag for human review.
9. Lower the temperature where the tool allows. It reduces variation, though as noted above it does not eliminate it.
10. Keep a record. Retain the prompt, version, date, inputs and output so decisions can be explained later to a lender, partner or adviser.

 

Biased and neutral prompt examples

These examples illustrate how a seemingly neutral prompt will create a biased result. It becomes misleading when possibilities turn into findings without new evidence, or a one-sided review is presented as the complete assessment. Research has documented both agreement with users’ views and reliance on early conversational assumptions.

 

1 Reviewing earnings adjustments

2 Reviewing compliance documents

3 Reviewing customer concentration

Bias and inconsistent answers are different problems

Bias concerns the direction of the answer. Non-determinism means identical inputs can produce different outputs. Changing the follow-up changes the input, so a changed answer alone does not prove non-determinism. Neutral wording reduces leading instructions; it cannot guarantee an unbiased or identical result.^[3]

If the conversation has drifted, restart from the source documents with the neutral prompt.

Do not carry earlier AI concerns forward as established facts. A fresh review still needs checking.

 

Biased and neutral prompts compared (summary)

A useful final discipline is the ‘reverse prompt’. Once the buyer has a draft conclusion, they ask the model to argue the opposite case as strongly as the evidence allows. If the counter-argument is weak, confidence is justified. If it is uncomfortable, it deserves attention before contracts are signed.

 

Treating risk neutrally

Buyers tend to make one of two errors with risk. Some minimise it because they want the deal. Others catastrophise it, treating every finding as a reason to walk away. Neither produces good decisions. A neutral approach treats risk as something to be described, measured, priced and managed.

The Australian Government’s Guidance for AI Adoption, released by the National AI Centre in October 2025, sets out six practices: decide who is accountable, understand impacts and plan accordingly, measure and manage risks, share essential information, test and monitor, and maintain human control.^[5] Although written for organisations adopting AI, those practices translate directly into a buyer’s use of AI in due diligence: a named person owns each conclusion, AI outputs are tested and a human makes the decision.

 

A neutral risk method

• Describe the risk factually. ‘Three of eight assessment tools rely solely on unsupervised written tasks’ rather than ‘assessment is a disaster’.
• Grade the evidence. Distinguish between risks confirmed by documents, risks indicated by inconsistencies and risks that are merely possible.
• Rate likelihood and consequence separately using defined scales, for example a 5 × 5 matrix aligned with ISO 31000.
• Identify the treatment. Can the risk be fixed before completion, after completion, through price or through contract terms?
• Assign a commercial response. Link each material risk to a specific mechanism: price, retention, warranty, indemnity, condition precedent or walk-away.
• Record what would change the rating. Note the document or fact that would raise or lower the risk, so the register stays live.

 

Using AI to identify opportunities

Due diligence is not only a search for problems. The same tools can help a buyer see value that a vendor has not captured. Three areas are particularly suited to AI-assisted analysis.

Scope and demand alignment. A buyer can compare a target’s scope with labour market and skills data to identify qualifications with strong demand that the RTO holds but under-uses, or adjacent skill sets that could be added. The NCVER data shows short-course enrolments increasing in fee-for-service markets while declining in government-funded training, which is relevant to how an RTO’s revenue mix might evolve.^[6]

Operational improvement. AI can help a buyer identify processes in the target that are manual, duplicated or error-prone: enrolment administration, student support enquiries, resource development, compliance self-assurance and reporting. These become the basis of a post-acquisition improvement plan. As discussed below, the value of that plan belongs to the buyer who executes it, not to the vendor who did not.

The caution applies here too. AI is prone to generating optimistic growth narratives. Every opportunity should be tested with the same discipline as every risk: what evidence supports it, what would it cost to realise, how long would it take and what regulatory approvals (such as scope additions) would be required?

 

How better AI efficiency could change the RTO opportunity

A new and important strategic question for buyers is not how AI helps them buy an RTO, but how AI changes what an RTO is worth. The answer depends on mechanisms that pull in different directions. Think about it; Ask how AI can increase efficiency in the RTO opportunity.

 

Forces that could increase value

• Lower cost to serve. Administration, enquiry handling, scheduling, resource development and internal audit preparation are labour-intensive in most RTOs. Jobs and Skills Australia identifies clerical and administrative roles as having the highest automation exposure.^[7] Where AI reduces those costs sustainably, maintainable earnings can rise.
• Efficiency gains allow an RTO to grow enrolments without proportional growth in overheads, which tends to improve margins and attract stronger buyer interest.
• Better self-assurance. AI tools that monitor data integrity, completion patterns and assessment consistency can support the continuous self-assurance the 2025 Standards emphasise, potentially lowering regulatory risk.
• New demand. The augmentation of existing jobs described by Jobs and Skills Australia creates demand for short courses and skill sets in digital and AI capability, and the Productivity Commission’s productivity outlook suggests employers will have reasons to fund that training.^[8],[9]

 

Forces that could reduce value

• Assessment integrity costs. Where student use of AI undermines written assessment, providers must redesign assessment and add verification such as observation or verbal questioning. That costs trainer time and may offset administrative savings. ASQA has made clear that integrity remains central to its expectations.^[10],[11]
• Competition erodes the gain. When every provider adopts the same tools, efficiency becomes the baseline rather than an advantage, and some of the saving is competed away through lower prices.
• Demand shifts in some qualifications. Qualifications aligned to highly automatable tasks may see softer long-term demand, which matters for scope-concentrated RTOs.
• Implementation risk. Australian SME adoption remains partial, at 43% in the National AI Centre’s December 2025 to February 2026 tracking.^[12] Projected savings that depend on technology the business has not yet implemented are forecasts, not earnings.

 

An illustrative example

The following figures are hypothetical and are included only to show the mechanics. Consider an RTO with revenue of $2.0 million and normalised maintainable earnings of $300,000 – 15%. A buyer believes AI-assisted enrolment administration, student support and resource development could save $80,000 a year after software costs, but also expects assessment redesign to add $30,000 a year in trainer time.

Three neutral observations follow.

First, the net gain is smaller than the headline saving.

Second, the $150,000 difference is a buyer’s opportunity that depends on the buyer’s capital, effort and execution risk. A buyer who pays the vendor for it in advance has transferred the reward while keeping the risk.

Third, for a vendor the position reverses: efficiencies that are already implemented, documented and reflected in at least 12 months of financial results can legitimately support maintainable earnings. Efficiencies that exist only as a plan generally cannot.

 

What this means for vendors and their advisers

Vendors who want AI to support their price should be able to show it, not describe it. That means a written AI use policy aligned with ASQA’s principles, evidence of how assessment authenticity is assured, a record of which processes use AI and the measurable effect on costs or service levels, and assurance that student personal information is handled consistently with privacy guidance. Consultants preparing an RTO for sale can treat these as part of standard sale readiness. Buyers will increasingly ask for them.

 

What not to do

• Do not upload confidential vendor material to public AI tools. Information memoranda and data rooms are usually covered by a confidentiality agreement. Check whether the tool retains or trains on inputs, and use business-grade products with appropriate contractual protections.
• Do not enter student personal information. The OAIC recommends, as best practice, that organisations not enter personal information, particularly sensitive information, into publicly available generative AI tools.^[13] Student records, USIs and complaint files should be de-identified or kept out entirely.
• Do not accept AI-generated facts about a specific RTO. Verify scope, registration history, CRICOS status and regulatory decisions directly from the public registers.
• Do not let AI set the price. A model has no access to comparable transaction data unless you supply it, and cannot carry professional responsibility for a valuation.
• Do not ask leading questions. Given the documented tendency of models to affirm their users, a leading prompt produces a confident echo of the buyer’s hopes.^[14]
• Do not rely on a single run or a single tool. Outputs vary.^[15] Material conclusions should be reproduced and checked by a person.
• Do not assume documents are genuine because they look professional. AI also makes forgery easier. In 2024 a finance worker at Arup in Hong Kong transferred HK$200 million across 15 transactions after a video call in which the ‘chief financial officer’ and colleagues were deepfakes.^[16] Verify bank statements, funding letters and identity documents at source.
• Do not treat AI output as legal, tax or accounting advice. It can help you prepare better questions for your advisers, not replace them.
• Do not overlook identity and AML obligations. Since 1 July 2026, lawyers, accountants and other professionals providing designated services, including assisting with the buying or selling of an entity, have AML/CTF obligations, so expect identity verification requests and plan for them.^[17]
  Do not ignore your own AI use after completion. Once you own the RTO, its AI policies and practices become your regulatory responsibility.

 

A note on common buyer mistakes

Many of the mistakes buyers make with AI are versions of mistakes buyers have always made;

• Falling in love with a business before diligence is complete
• Anchoring on the asking price
• Under-resourcing compliance review and confusing a vendor’s narrative with evidence.

AI does not create these errors, but it can make them faster and more convincing. They deserve fuller treatment than space allows here and will be the subject of a companion article.

 

Conclusion

AI is changing how RTO acquisitions are researched and prepared. The evidence suggests it can make a careful buyer faster and more thorough. The same evidence suggests it can make a careless buyer more confident without making them more correct. The difference lies almost entirely in method: neutral questions, primary sources, structured outputs, repeated runs and human accountability.

For buyers, the opportunity is twofold. AI can improve the quality of the purchase decision, and it may improve the economics of the business purchased. For vendors, the lesson is that AI-related value must be demonstrated in results and governance, not asserted. For consultants, AI is becoming part of sale readiness and due diligence alike. In a sector where the regulator’s focus is integrity, the buyers who benefit most from AI will be those who use it with the most integrity.

 

 

 

About the author. Travis Latter is CEO of Infinity Business Brokers and RTO Valuations, which specialise in the sale, acquisition and valuation of Australian registered training organisations. Disclosure: as a broker and valuer, the author’s firm acts for parties in RTO transactions. This article is general information only and is not legal, financial, tax or valuation advice. Readers should obtain advice specific to their circumstances. The research and drafting of this article were assisted by AI tools; all sources were reviewed and the content approved by the author.

 

 

[1]ASQA, ‘2025 Standards for RTOs commence’, 1 July 2025; ASQA, Regulatory Risk Priorities 2025–26, including ‘Integrity of qualifications and competency outcomes’ and ‘Transparency, accountability and assurance in complex and international delivery models’ (asqa.gov.au).

[2]ASQA, Regulation Report Q1–2 2025–26 (published March 2026) and Regulation Report Q1–3 2025–26 (published June 2026) (asqa.gov.au).

[3]Bain & Company, M&A Report 2026, ‘Capability for a new era’ (survey of more than 300 M&A executives) (bain.com).

[4]ASQA, Principles for the Responsible Use of AI in VET (2026), as summarised in Training Accreditation Council WA, TAC Update Issue 123, October 2026 (wa.gov.au). ASQA states the principles do not create new regulatory requirements.

[5]National AI Centre, Guidance for AI Adoption: Foundations and Implementation Practices, October 2025 (industry.gov.au; ai.gov.au). The guidance replaced the 2024 Voluntary AI Safety Standard.

[6]NCVER, Total VET students and courses 2025, research summary, 22 July 2026 (ncver.edu.au).

[7]Jobs and Skills Australia, Our Gen AI Transition: Implications for Work and Skills, Overarching Report, 14 August 2025 (jobsandskills.gov.au); modelling results as reported by Business Insider, August 2025.

[8]Jobs and Skills Australia, Our Gen AI Transition: Implications for Work and Skills, Overarching Report, 14 August 2025 (jobsandskills.gov.au); modelling results as reported by Business Insider, August 2025.

[9]Productivity Commission, Harnessing Data and Digital Technology, final report, 19 December 2025, as summarised by King & Wood Mallesons, December 2025.

[10]ASQA, Principles for the Responsible Use of AI in VET (2026), as summarised in Training Accreditation Council WA, TAC Update Issue 123, October 2026 (wa.gov.au). ASQA states the principles do not create new regulatory requirements.

[11]ASQA, ‘2025 Standards for RTOs commence’, 1 July 2025; ASQA, Regulatory Risk Priorities 2025–26, including ‘Integrity of qualifications and competency outcomes’ and ‘Transparency, accountability and assurance in complex and international delivery models’ (asqa.gov.au).

[12]National AI Centre, ‘AI adoption insights: December 2025 to February 2026’, first published 7 May 2026 (ai.gov.au).

[13]Office of the Australian Information Commissioner, Guidance on privacy and the use of commercially available AI products, 21 October 2024 (oaic.gov.au).

[14]Cheng M, Lee C, Khadpe P, Yu S, Han D and Jurafsky D, ‘Sycophantic AI decreases prosocial intentions and promotes dependence’, Science, March 2026.

[15]He H and Thinking Machines Lab, ‘Defeating Nondeterminism in LLM Inference’, Connectionism, September 2025 (thinkingmachines.ai).

[16]CNN, ‘British engineering giant Arup revealed as $25 million deepfake scam victim’, May 2024; Hong Kong Police Force briefing reported February 2024.

[17]Law Society of NSW, ‘Understanding designated services: when legal services trigger Tranche 2 AML/CTF obligations’, 2026; Corrs Chambers Westgarth, ‘Navigating the AML/CTF reforms: extending the regime to Tranche 2 entities’, 2025.

 

Appendix: A practical checklist

• Use a business-grade AI tool with confirmed data protections, and check the confidentiality agreement before uploading anything.
• Write and version-control your screening, review and risk prompts before you start.
• Pull all RTO-specific facts from primary registers and paste them in; never ask the AI to supply them.
• Require citations to the documents provided and permit ‘Not evidenced’ as an answer.
• Separate facts, assumptions and opinions in every output.
• Run material prompts at least twice and investigate differences.
• Run a reverse prompt against every draft conclusion.
• Grade every risk and every opportunity by the strength of evidence.
• Map each material risk to a commercial response.
• Have a named person, and your professional advisers, own every conclusion that affects price or structure.

 

Appendix: Starter prompt library

The prompts below are written to be neutral and repeatable. Replace bracketed text, keep a version number and use them only with tools approved for confidential material.

Prompt 1 — Screening (v1.0)

Role: You are an independent reviewer screening an RTO acquisition opportunity. You have no stake in the outcome.

Inputs: [Information memorandum text] and [extract from training.gov.au].

Task: Score the opportunity against each criterion in the attached scoring table using only the inputs. For each criterion give the score, the supporting page or section and a confidence rating (High/Medium/Low).

Rules: If a criterion is not addressed, write ‘Not evidenced’ and score it as blank. Do not estimate. List any conflicts between the memorandum and the register extract.

Output: A table with columns Criterion | Score | Evidence reference | Confidence | Notes.

 

Prompt 2 — Balanced assessment (v1.0)

Using only the attached documents, set out the five strongest evidence-based reasons to acquire this RTO and the five strongest evidence-based reasons not to.

Give each side equal depth. Cite the document and section for every point.

Then list the three facts that, if confirmed, would most change the overall assessment in either direction.

 

Prompt 3 — Data room gap analysis (v1.0)

Compare the attached data room index with the attached due diligence checklist.

Produce a table: Checklist item | Matching document(s) | Status (Present / Partial / Missing) | Reason.

For every Partial or Missing item, draft a short, neutral information request to the vendor. Do not assume a document’s content from its file name; mark such items ‘Unverified’.

 

Prompt 4 — Normalisation challenge (v1.0)

Review the vendor’s proposed add-backs against the attached financial statements and supporting documents.

Classify each as: Supported by documents provided / Plausible but unsupported / Not supported. Give a one-line reason and the document that would resolve any uncertainty.

Separate your output into Facts, Assumptions and Open questions. Do not calculate a valuation.

 

Prompt 5 — Neutral risk register (v1.0)

From the attached review notes, build a risk register using the attached likelihood and consequence scales.

For each risk record: description (factual, no adjectives) | evidence grade (Confirmed / Indicated / Possible / Not assessed) | likelihood | consequence | treatment | possible commercial response | what evidence would change the rating.

Also record an equal number of opportunities in the same format. Do not soften or exaggerate; describe only what the evidence supports.

 

Prompt 6 — Reverse test (v1.0)

Here is my draft conclusion: [conclusion]. Argue the opposite conclusion as strongly as the attached evidence genuinely allows.

Identify any point where my draft relies on an assumption rather than evidence.

Finish with a plain statement of whether the counter-argument is weak, moderate or strong, and why.