AI Governance for Financial Advice and Accounting Practices
The government's November 2026 AI rules are aimed at the companies that build AI. A practice that uses it already answers to ASIC, the OAIC, AUSTRAC and the Tax Practitioners Board. Who regulates what, which uses carry the most risk, and a five-stage method for getting AI under control, with a 20-step checklist.
The short version
On 10 November 2026 the government accepted all 18 recommendations of its Rapid Review into the OpenAI agent incident at Services Australia, and announced mandatory AI safety rules with a new regulator to enforce them. Those rules are aimed at the companies that build frontier AI. They are not yet law, and they do not regulate a practice that uses AI.
That does not leave an advice or accounting practice unregulated. Your use of AI is already covered by rules you answer to today:
- ASIC, through your licensee, for financial advice. Existing obligations apply to AI as they do to anything else.
- The OAIC, for client information, including the Privacy Act's new automated decision duty from 10 December 2026.
- AUSTRAC, if you are a reporting entity, which many accountants became on 1 July 2026.
- The Tax Practitioners Board, whose AI guidance for tax and BAS agents was issued on 22 July 2026.
This guide sets out a five-stage method for getting AI under control in a small practice: find what you use, rate it, put ten controls in place, roll tools out in waves, and review. It ends with a 20-step checklist and a register template.
- Who this is for
- What changed in November 2026
- Who regulates what
- Common AI uses, rated by risk
- Find what you already use
- What Microsoft 365 already shows you
- Ten minimum controls
- The automated decision check
- Questions for AI vendors
- When an AI tool goes wrong
- The five stages, as a plan
- The AI register
- 20-step checklist
- Questions
Who this is for
You are probably not regulated by APRA. APRA supervises banks, insurers and most super funds. A practice operating under someone else's AFS licence is not one of those, so CPS 230 and CPS 234 do not bind you directly. They can reach you through your licensee, if it belongs to a banking, insurance or super group and models its policies on them. Treat them as a benchmark, and see our AI under APRA CPS 230 and 234 guide if you want one.
If you are an authorised representative, start with your licensee. The licensee is responsible to ASIC for what you do under its licence, and most now have an AI policy and a list of approved tools. Ask for both before you choose anything. Your own rules have to fit inside theirs.
What changed in November 2026
On 18 June 2026 an OpenAI research agent was refused by the Medicare Statistics Reporting Service portal, found another way in, read files that were not public and wrote files to a Services Australia server. The government was told 84 days later, by an email to a public mailbox. Our analysis of the incident is here.
Two announcements followed on 10 November:
- The Rapid Review's 18 recommendations, all accepted. The review was led by the Department of the Prime Minister and Cabinet with the National Cyber Security Coordinator, the Office of AI, the Australian Signals Directorate, the AI Safety Institute and Services Australia.
- Mandatory AI safety and security protections, to be legislated through the Australian AI Standards and enforced by a new AI safety regulator working with the AI Safety Institute. They apply to "all frontier AI companies that wish to operate in Australia".
Frontier AI companies would have to keep security controls across development and deployment, identify and manage serious risks, allow independent evaluation of their models, notify the government of incidents rapidly, publish their safety and security frameworks, and respond as the standards change. The regulator would be able to impose significant penalties. None of this is law yet, and the government says more detail will follow consultation.
The 18 recommendations, and what they mean for a practice
| Theme | Recommendations | What the review asks for | What it means for you |
|---|---|---|---|
| Reporting by AI developers | 3, 4 | Consider mandatory, time-bound AI safety reporting, covering misaligned behaviour and agentic cyber incidents. Explore protections for those who report incidents and near misses. | Your AI vendors may have to tell the government. Nothing yet makes them tell you. Put a notification clause with a number of hours in every AI contract. |
| Criminal law | 8 | Consider criminal penalties for egregious regulatory non-compliance, and a standalone "failure to prevent" offence for a company whose model causes harm when it did not take reasonable steps to prevent it. | Aimed at developers, but it shows where the test is heading: reasonable steps, and evidence of them. Keep a record of the steps your practice took. |
| Standards and engagement | 1, 2, 5, 6, 7 | Closer engagement with frontier AI developers, AI Standards built on accepted international approaches, and a larger role for the AI Safety Institute. | The AI Standards legislation will cover more than frontier safety. The six AI6 practices are the best preparation for whatever it requires of AI users. |
| Escalating incidents | 9, 10, 11, 12, 13 | Clearer rules for agencies on escalating reported vulnerabilities and suspected incidents before they are technically confirmed, better information sharing, and crisis plans that account for AI by the end of 2026. | The warning in this incident reached a general inbox. Check where a vendor's warning would land in your practice, and who would act on it. |
| Cyber security basics | 14, 15, 16, 17, 18 | Lift AI literacy among staff, return to the fundamentals including ASD's Essentials, shut down or segregate systems that do not need the internet, harden critical systems, and use AI-enabled defence responsibly. | The same list works for a practice: train staff, cover the basics, and switch off old systems and portals nobody uses. |
The recommendations are grouped here by CISO AI. The government published them as a single list; the full wording is on the Department of the Prime Minister and Cabinet's website, linked under Sources.
Who regulates what
There is still no single AI law for businesses that use AI. A practice answers to the regulators it already has, each for a different part of the same client file.
| Regulator | Covers | What it says about AI | What to do |
|---|---|---|---|
| ASIC, through your licensee | Financial advice and other financial services | Report 798 (October 2024) found licensees adopting AI faster than their governance. ASIC's position is that existing obligations are technology neutral: acting efficiently, honestly and fairly, the best interests duty and advice record keeping all apply to AI-assisted work. The report includes 11 questions for licensees. | Get your licensee's AI policy. Use the 11 questions as a self-check. An adviser stays responsible for advice a tool drafted. |
| OAIC | Personal information | The Australian Privacy Principles apply to what goes into an AI tool and what comes out. APP 8 covers sending information overseas, APP 11 covers security, and the Notifiable Data Breaches scheme covers leaks. From 10 December 2026, privacy policies must disclose significant automated decisions. | Run the automated decision check. Know where each tool stores data. Confirm whether the small business exemption applies to you (see below). |
| AUSTRAC | Reporting entities under the AML/CTF Act | Accountants that provide designated services became reporting entities on 1 July 2026. If AI is used for identity checks, client risk rating or monitoring, the AML/CTF program must describe it, and the compliance officer remains accountable for the result. | Write AI use into the program. Keep suspicious matter information out of general-purpose AI tools: the tipping-off offence restricts who can see it. |
| Tax Practitioners Board | Registered tax agents and BAS agents | TPB(GS) 55/2026, issued 22 July 2026, applies the Code of Professional Conduct to AI. It adds no new rules: the practitioner remains accountable, must take reasonable care, must stay competent, and must keep client information confidential. | Review AI output before it is relied on. Cover AI tools and offshore processing in engagement letters. Keep a record of the review. |
| Professional bodies | Members of CA ANZ, CPA Australia and the IPA | The APES 110 Code of Ethics duties of confidentiality and professional competence apply whatever tool produced the work. | Treat AI output like the work of an unsupervised junior: check it before it leaves. |
| The proposed AI safety regulator | Frontier AI companies | Announced 10 November 2026, not yet legislated. No obligations on businesses that use AI. | Watch the AI Standards consultation. Ask vendors which model sits under their product. |
The small business exemption is narrower than it looks
Businesses with annual turnover of $3 million or less are generally exempt from the Privacy Act. Three things cut into that for a practice:
- Tax file numbers. The Privacy (Tax File Number) Rule binds anyone who holds TFNs, whatever their size, and a TFN breach can be notifiable.
- AML/CTF. A small business that is a reporting entity is covered by the Privacy Act for its AML/CTF activities, though not necessarily for everything else it does.
- Your licensee and your clients. Licensee agreements and client contracts often require you to meet the Privacy Principles regardless.
Work out which of your activities are covered before you decide what the 10 December change requires. The government has also agreed in principle to remove the exemption in later reforms, so a practice that builds to the Privacy Principles now will not have to redo the work.
Common AI uses, rated by risk
The rating is a starting point for your own register. It reflects what happens if the tool is wrong or the data leaks, and which regulator would care. Red does not mean banned. It means the use needs sign-off, a human check on every output, and usually your licensee's approval.
| Use | Rating | Why | Before it goes live |
|---|---|---|---|
| Drafting marketing, internal documents and templates with no client data | Green | No personal information, and a person reads it before it is used. | An approved tool and the acceptable use policy. |
| Summarising public material: rulings, legislation, product disclosure statements | Green | Public inputs. The risk is a wrong or invented citation. | Check every citation against the source. |
| Meeting recording and file notes | Amber | Records sensitive financial and health details. The file note is part of the advice record, and transcription errors become the record. | Client consent to record, a vendor check on storage and training use, and the adviser reviews each note. |
| Drafting client emails and letters | Amber | Client data in the prompt, and a risk that general information drifts into personal advice. | An approved tool only, and a person sends it. |
| Bookkeeping automation: coding transactions, bank reconciliation, extracting data from documents | Amber | Errors flow into BAS and tax returns. Source documents often carry TFNs and bank details. | Sampling of the output, and a check on where documents are processed. |
| Tax and technical research with client facts | Amber | The TPB expects competence and reasonable care. AI can state the law wrongly with confidence. | Remove identifying details where possible, and verify against primary sources. |
| Drafting statements of advice and records of advice | Red | The best interests duty and the licensee's obligations attach to the document. The adviser, not the tool, is responsible for it. | Licensee approval, adviser review of every document, and the draft and the review kept on file. |
| Risk profiling and product or strategy recommendations | Red | A program that shapes the advice a client receives is likely to be an automated decision under the Privacy Act. | The automated decision check, licensee approval, and an adviser who can override the result. |
| Identity verification and AML risk rating | Red | The result can lead to enhanced checks or a refusal to act. AUSTRAC holds the compliance officer accountable for it. | The AML/CTF program describes the tool, a person reviews high-risk and refused cases, and the automated decision check is done. |
| Preparing and lodging tax returns and BAS | Red | The registered agent is accountable for every lodgment under the Code of Professional Conduct. | The agent reviews before lodgment, and the review is recorded. |
| AI agents that act on their own: logging in to portals, moving data between systems, sending messages | Red | The June incident was an agent that did not accept a refusal. An agent using your credentials acts as you. | Its own account with the least access it needs, approval before it acts, and logging. See our securing AI agents guide. |
| Client-facing chatbots | Red | An answer tailored to a client's circumstances can amount to personal advice given without an adviser. | Limit it to factual and administrative answers, tell clients it is AI, and hand over to a person. |
Find what you already use
Most practices use more AI than they think. It arrives as a feature switched on inside software they already pay for, or as a tool one staff member signed up to. Discovery draws on four sources, because no single one gives the whole picture.
- A short questionnaire
- Which systems the practice runs (practice management, CRM, advice software, accounting ledger, AML or identity tool, Microsoft 365 or Google Workspace), which AI tools and AI features are in use, which are being considered, and which Microsoft 365 plan the practice is on.
- A document request
- The privacy policy and collection notices, the AML/CTF program, IT and security policies, client engagement letters, key vendor contracts, and the licensee's AI policy and approved tool list.
- Conversations with staff
- Fifteen minutes each with an adviser, an accountant, an administrator and whoever looks after IT. Ask what they use, including on their own phone, and what they wish they could use. People answer honestly when the aim is to approve tools, not to catch anyone.
- A read-only technical check
- What the practice's own systems record about AI use. For a practice on Microsoft 365, this needs no new software.
What Microsoft 365 already shows you
What you can see depends on the licence. An outside reviewer needs only a read-only role, such as Global Reader or Security Reader, and that access should be written into the engagement letter.
| Where to look | Licence | What it shows |
|---|---|---|
| Entra ID: enterprise applications and user consent settings | Every plan | Third-party apps staff have connected to their work accounts, including AI note-takers and add-ins, and what each app is allowed to read. You can also stop staff approving new apps without an administrator. |
| Copilot usage reports and the audit log | Every plan with Copilot | Who uses Microsoft 365 Copilot and how much. |
| Defender for Cloud Apps: Cloud Discovery, filtered to generative AI | Depends on the plan. Sources differ on what Business Premium includes, so check the tenant. | The AI services staff actually visit from enrolled devices, with users, the volume of data uploaded and a risk score for each service. |
| Purview: Data Security Posture Management for AI and endpoint data loss prevention | E5, or an add-on to other plans | What kinds of sensitive information go into AI tools, with the option to warn or block when someone pastes a TFN or a client file. |
What it will not show you
- Personal phones and home computers that are not enrolled.
- AI features inside software you already use, such as advice software, the ledger or the CRM. Discovery shows the product is in use, not that its AI is switched on.
- History. Discovery data is kept for a limited period, around 90 days by default, so it is a snapshot.
That is why the questionnaire and the conversations matter as much as the scan. Seeing AI use and blocking it are also different things: blocking usually needs a higher licence, and it comes after the practice has decided what is approved.
Ten minimum controls
These are sized for a practice of five to fifty people. Each maps to one of the six practices in the government's Guidance for AI Adoption (AI6), the government's main guidance for organisations using AI. Our AI guardrails guide covers AI6 in full.
| Control | What it looks like in a small practice | AI6 practice |
|---|---|---|
| 1. An accountable owner | One principal or director is named as responsible for AI. Each tool also has an owner who knows how it is used. | 1. Decide who is accountable |
| 2. An AI register | One sheet listing every tool and AI feature, what it does, the data it touches, where that data goes, its rating and its owner. | 2. Understand impacts |
| 3. An acceptable use policy | Two pages: the approved tools, what each may be used for, and what never goes into an unapproved tool, including TFNs, identity documents, client files and suspicious matter information. | 1. Decide who is accountable |
| 4. A rating and sign-off before any new use | Green, amber or red, with the reason. Amber needs the tool owner's approval. Red needs the accountable owner and, for advice, the licensee. | 3. Measure and manage risks |
| 5. Human review of advice and lodgments | A qualified person reads and approves anything AI drafted before it reaches a client, the ATO or another regulator, and can overrule the tool. | 6. Maintain human control |
| 6. Vendor due diligence | Every AI vendor answers the questions below before a contract is signed, and again when the product changes. | 3. Measure and manage risks |
| 7. Rules for client data | Which kinds of client information each tool may hold, whether it is stored or processed overseas, and whether the engagement letter allows that. | 3. Measure and manage risks |
| 8. Telling clients | The privacy policy and engagement letters say how AI is used. Clients are asked before a meeting is recorded. Automated decisions are disclosed. | 4. Share essential information |
| 9. Records and monitoring | The AI output, the review and the approval are kept on the client file. A sample of each tool's output is checked every quarter. | 5. Test and monitor |
| 10. Incident handling and training | Staff know what to do when a tool is wrong or data goes where it should not, and are trained on the policy before they get access. | 5. Test and monitor |
The automated decision check
From this date, privacy policies must disclose significant automated decisions. The duty is in Australian Privacy Principles 1.7 to 1.9, and the OAIC published a fact sheet, a flowchart and updated APP 1 guidelines on 30 September 2026.
The duty applies to a use only if all three of these are true:
- A computer program makes the decision, or does something substantially and directly related to making it. This is not limited to AI: a scoring rule in a spreadsheet counts.
- The decision could reasonably be expected to significantly affect the rights or interests of an individual.
- Personal information about that individual is used by the program.
How common uses are likely to fall
| Use | Likely result | Reasoning |
|---|---|---|
| A tool rates a new client's money laundering risk, and the rating decides whether you act for them | Disclose | A program substantially informs a decision to provide or refuse a service, using identity and financial details. |
| A risk profiling tool sets a client's risk category, which drives the recommended portfolio | Disclose | The category directly shapes the advice, even though an adviser signs it. |
| Software pre-assesses a client's eligibility for a loan or an insurance product | Disclose | Access to credit or cover is a significant effect. |
| AI drafts a statement of advice that the adviser rewrites and approves | Assess | It turns on how much the adviser changes. A draft that is approved largely unread substantially informs the decision. |
| AI transcribes a meeting and writes the file note | Not a decision | No decision is made. Recording consent and the Privacy Principles still apply. |
| Bookkeeping software codes transactions | Not a decision | No significant effect on an individual's rights or interests. |
Record your reasoning for each use, including the ones you decide are out of scope. Using a vendor's product does not move the duty to the vendor: the practice that arranged for the program to be used is the one that must disclose it.
Example wording for a privacy policy
Adapt the wording to what your practice does. The policy must name the kinds of personal information used, the kinds of decisions made solely by a program, and the kinds a program substantially and directly informs.
Questions for AI vendors
Ask these of every AI tool and every AI feature inside existing software. A vendor that cannot answer them is giving you an answer.
Your data
- Where is our data stored and processed, and does any of it leave Australia?
- Is our data, or our clients' data, used to train or improve any model? Is that off by default, and is it in the contract?
- How long are prompts, uploads, recordings and outputs kept, and can we delete them?
- Which other companies handle our data, and which AI model sits under your product?
- When we leave, how do we get our data out, and when is your copy deleted?
Security
- Does the product support single sign-on and multi-factor authentication, and does it keep a log of who did what?
- Which independent assurance do you hold: ISO 27001, SOC 2, ISO 42001?
- Does the product take actions on its own, such as sending messages or signing in to other systems? What stops it when it is refused?
When something goes wrong
- Within how many hours will you tell us about an incident that affects our data or our systems? The answer should be a number.
- Will you tell us when you suspect an incident, or only once you have confirmed it?
- Who will you tell? Give the vendor a named contact and a monitored address, not a general inbox.
- What do you claim about accuracy, and what happens when the output is wrong?
The last group is the lesson of 2026. The government learned of the Services Australia incident 84 days after it happened. "Prompt" and "without undue delay" did not produce a phone call.
When an AI tool goes wrong
Decide in advance who makes the call and who has to be told. Four situations cover most of what a practice will meet.
| What happened | First steps | Who may need to know |
|---|---|---|
| Client data went into a tool it should not have, or a vendor was breached | Stop the use, find out what was exposed, and ask the vendor to delete it. | Assess it as a data breach: an eligible breach must be notified to the OAIC and the people affected, and the assessment should take no more than 30 days. Tell your licensee. |
| Wrong AI output reached a client, the ATO or a product provider | Correct it with the recipient, and review other work the tool produced in the same way. | Your licensee, which has its own breach reporting duties to ASIC. The TPB, where a tax practitioner's breach of the Code is significant. Your professional indemnity insurer. |
| A tool took an action it should not have | Disable it, revoke its access and passwords, and keep the logs. | The owner of any system it touched, and your licensee. |
| A vendor tells you about an incident on their side | Ask what data and which dates, in writing, and treat it as your incident until you know otherwise. | As above, depending on what was affected. |
The five stages, as a plan
An example for a practice of about twenty people. The timing assumes a few hours a week from one principal and one administrator.
| Stage | When | What happens | What you have at the end |
|---|---|---|---|
| 1. Discover | Weeks 1 to 2 | The questionnaire, the document request, staff conversations and the read-only technical check. | A first AI register, and the licensee's AI policy in hand. |
| 2. Assess | Weeks 3 to 4 | Rate each use green, amber or red. Run the automated decision check. Send the vendor questions. Compare what you found with the regulator map. | A rated register and a short list of gaps, in priority order. |
| 3. Govern | Weeks 4 to 6 | Name the owner. Write the acceptable use policy. Update the privacy policy, engagement letters and the AML/CTF program. Set the review and incident steps. | The ten controls, written down and approved. |
| 4. Roll out | Weeks 6 to 12 | Introduce tools in three waves, each with its own gate (below). Train staff before each wave. | Approved tools in use, with a record of who approved what. |
| 5. Review | Every quarter | Recheck the register, sample each tool's output, repeat the technical check, and pick up changes in the rules. | A register that is still true, and evidence of reasonable steps. |
The three waves, and the gate before each
| Wave | Tools | In place before it starts |
|---|---|---|
| Wave 1 | Green Drafting and research with no client data | The owner, the register, the acceptable use policy and staff training. |
| Wave 2 | Amber Meeting notes, client correspondence, bookkeeping automation | Vendor answers on file, engagement letters and the privacy policy updated, recording consent in the meeting process, and a review step for each tool. |
| Wave 3 | Red Advice documents, risk profiling, AML checks, lodgments, agents | Licensee approval, the automated decision disclosure published, the AML/CTF program updated, the human review procedure tested on real work, and the incident steps rehearsed. |
The AI register
One row for each tool or AI feature. The same sheet serves the licensee, the privacy policy review and the quarterly check, so it is worth keeping accurate.
| Column | What goes in it |
|---|---|
| Tool and vendor | The product, the vendor, and the AI model underneath if known. |
| What it is used for | The task, and which team uses it. |
| Client data it touches | None, contact details, financial information, health information, TFNs, identity documents. |
| Where the data goes | Stored and processed in Australia or overseas, kept for how long, used for training or not. |
| Who checks the output | The role that reviews it, and whether every output or a sample. |
| Rating | Green, amber or red, with the reason. |
| Automated decision? | Yes, no or assess, against the three-part test, with the reasoning. |
| Approvals | Tool owner, accountable owner, licensee where needed, and the date. |
| Next review | The date, and what would trigger an earlier one. |
Download the register template (CSV). It opens in Excel or Google Sheets and includes two example rows.
AI governance checklist for a practice
Twenty checks across the five stages. Ticks are stored in this browser only.
Discover: AI6 practice 1
Discover: ASIC REP 798
Discover: AI6 practice 2
Discover: Microsoft 365 and staff conversations
Assess: AI6 practice 3
Assess: APP 1.7 to 1.9
Assess: small business exemption
Assess: APP 8 and APP 11
Govern: AI6 practice 1
Govern: TFN Rule; AML/CTF Act
Govern: AI6 practice 6; TPB(GS) 55/2026
Govern: APP 1.7 to 1.9, from 10 Dec 2026
Govern: TPB Code; APP 8
Govern: privacy and surveillance devices law
Govern: AML/CTF Act
Govern: ASIC and TPB record keeping
Roll out: AI6 practice 1
Roll out: vendor contracts
Roll out: NDB scheme; licensee; TPB
Review: AI6 practice 5
Want this run for your practice?
Cintelis, the company behind CISO AI, runs the five stages in this guide as an AI Assurance & Advisory engagement for advice and accounting practices, starting with a read-only discovery. Tell us where you are and we will reply within two business days.
CISO AI is published by Cintelis Pty Limited, which delivers this service. Your details are used only for this enquiry; see our privacy policy.
Questions
Do the new AI safety rules apply to my practice?
No. The mandatory protections announced on 10 November 2026 would apply to frontier AI companies, meaning the developers of the most advanced models, and they are not yet law. A practice that uses AI is governed by the rules it already answers to: ASIC through its licensee, the Privacy Act, the AML/CTF Act and the Tax Practitioners Board's Code of Professional Conduct.
Is a financial advice or accounting practice regulated by APRA?
Almost never. APRA supervises banks, insurers and most superannuation funds. An authorised representative or an accounting practice is not one of those. APRA's standards can still reach a practice indirectly, through a licensee that belongs to an APRA-regulated group.
We are under $3 million turnover. Does the Privacy Act apply to us?
Possibly in part. The small business exemption does not cover tax file numbers, which are protected whatever your size, and a small business that is an AML/CTF reporting entity is covered for its AML/CTF activities. Licensee agreements and client contracts often require compliance as well. Work out which activities are covered before deciding what the 10 December 2026 change requires.
Does an AI meeting note-taker count as automated decision-making?
Usually not, because it records and summarises without making a decision about anyone. It still handles sensitive client information, so you need the client's consent to record, a check on where the vendor stores the recording and whether it trains on it, and an adviser who reviews each note.
Can we find shadow AI without buying new software?
Largely, yes, if the practice uses Microsoft 365. Every plan shows the third-party apps staff have connected to their work accounts, and some plans show which AI services are visited from enrolled devices. It will not show personal devices or AI features inside other software, which is why a questionnaire and staff conversations are still needed.
Who is responsible when AI gets advice or a tax return wrong?
The practitioner. ASIC treats existing obligations as technology neutral, so the adviser and the licensee remain responsible for advice a tool drafted. The Tax Practitioners Board's guidance says a registered agent remains accountable for work produced with AI.