AI Inquiry Atlas

An unofficial, independent guide to the public records of the Joint Select Committee on Artificial Intelligence. Not connected with the Parliament of Australia or the committee. Its makers also made a submission (no. 481; see About). Summaries and scores are written by AI and can be wrong: check the original before relying on them.

Submission 555

Ms Jessica Browne

Submitter group: Individuals.

Position labels are not shown for individual submitters.

An AI model identified the submitter group and wrote the summary, key points and recommendation summaries. Counts do not measure importance or merit.

Open in the interactive guide

Summary

Read the original submission for the full text.

A registered tax practitioner argues, using accounting software as an example, that AI embedded in connected financial systems creates systemic economic resilience risks and accountability gaps, while stating support for AI's productivity opportunities. Makes twenty numbered recommendations and additional requests on dependency mapping, evidence requirements, segregation of agentic authority, backup, supply chain disclosure, data sovereignty and stress testing.

Key points

Quotes are the submitter’s own words. Text taken from documents, including scanned pages read by text-recognition software, can differ from the PDF. Check the original. Listed in the same order as on the submission’s page in the interactive guide.

  1. States that the submission is not principally about tax practitioners but about the resilience of the Australian economy.

    This submission is not principally about tax practitioners. It is about the resilience of the Australian economy.

  2. Argues that an error in technology used across many businesses, flowing automatically into connected systems, could become a correlated economic failure.

    If the same error is created by technology used across a very large number of Australian businesses, and that error can flow automatically into other connected systems, we have a very different problem. That is no longer simply an accounting error. It can become a correlated economic failure.

  3. Describes 'Impossible Compliance', where a person is held accountable without the information, evidence or technical control to meet that responsibility.

    I call it Impossible Compliance. It occurs when a person is held accountable for something but does not have the information, evidence or technical control necessary to meet that responsibility.

  4. Argues Australian hosting does not resolve processors’ legal jurisdiction, while expressly stating this does not mean business data can simply be accessed without legal process.

    My point is not that Australian business data can simply be accessed without legal process. It is that the legal jurisdiction of the entities processing Australian data remains relevant even where the primary data is described as hosted in Australia.

  5. Says small businesses may not consciously adopt AI because it can arrive through updates to software they already use.

    small businesses may not consciously decide to adopt AI. AI can simply arrive inside software they already use.

  6. Argues that AI removing experience-building work raises questions about how future professionals will develop judgement to supervise autonomous systems.

    If AI progressively removes the work through which people acquire that experience, Australia needs to consider how the experienced professionals expected to supervise increasingly autonomous systems in ten or twenty years will develop their judgement.

Recommendations

Recommendations are shown beside the submission’s own words. Listed in the same order as on its page in the interactive guide.

  1. Assess productivity beyond removed hours or jobs, accounting for error correction, cybersecurity, verification, insurance, records, continuity, disputes and responsibility remaining with practitioners.

    However, I do not think Australia should measure the productivity benefit simply by counting the hours or jobs removed. Someone still has to deal with incorrect outputs, cybersecurity, verification, insurance, record keeping, business continuity and disputes. In a regulated profession, the work may disappear while the responsibility does not. A practitioner may spend less time preparing something but more time determining whether it can safely be relied upon.

  2. Assess national productivity and where automation moves economic value, replacement jobs, intellectual property and tax income; this does not mean preventing automation.

    There is also a difference between a productivity gain for one business and a productivity gain for Australia. If Australian wages are replaced with subscriptions to foreign software, cloud infrastructure, AI models and APIs, government should consider where the economic value has moved, where the replacement jobs are being created, where the intellectual property sits and where the resulting income is taxed. This does not mean automation should be prevented. It means productivity should be considered across the whole economy.

  3. Consider whether ordinary taxpayers and microbusinesses can realistically provide professional-equivalent supervision; self-service should not transfer professional and technology risk to those least able to assess it.

    If registered professionals are expected to understand, challenge and document AI-assisted work, government should consider whether it is realistic to expect an ordinary taxpayer or microbusiness owner to provide equivalent supervision. Self-service should not quietly transfer professional and technology risk to the person least able to assess it.

  4. Treat material autonomous-capability, permission, data-access and third-party-processing changes differently from ordinary updates; provide plain information on AI operation, access, changes, autonomy, third parties and available evidence.

    I believe material changes in autonomous capability, permissions, data access and third-party processing should be treated differently from ordinary software updates. A new colour scheme is an ordinary update. Giving software new authority to read, reason over, alter or act upon financial information is not the same thing. A sole practitioner cannot reasonably be expected to be an accountant, tax specialist, cybersecurity engineer, privacy specialist, lawyer, AI auditor and cloud architect at the same time. Businesses need plain information about what AI is operating, what information it can access, what it can change, whether it can act autonomously, what third parties are involved and what evidence is available if something goes wrong.

  5. Understand foreign-controlled financial-system infrastructure’s jurisdiction, concentration, exit options and consequences of access, terms, capability or availability changes; preserve practical operation, evidence, recovery and provider-switching capability.

    Where essential Australian financial activity depends on foreign-controlled cloud, AI, identity, communications and software infrastructure, government should understand the legal jurisdiction, concentration risk, exit options and practical consequences if access, terms, capability or availability change. Sovereignty requires more than Australian hosting. It requires the practical ability to operate, preserve evidence, recover and change providers without losing control of the underlying records.

  6. Make clear whether business and professional information serves delivery, training, testing, personalisation, product development, behavioural analysis or other AI purposes.

    I believe businesses and professionals should know how this information is being used. It should be clear whether information is being used simply to provide the service or for training, testing, personalisation, product development, behavioural analysis or other AI purposes. Saying that data is "owned" by the customer does not necessarily answer those questions.

  7. Map the full data and authority chain: processors, incorporation, subprocessors, AI providers, derived data, metadata, applicable law, automated authority and independently retained evidence, rather than only database location.

    In my view, government should therefore map the full data and authority chain, not merely the location of the primary database. That means understanding which entities can process the information, where those entities are incorporated, what subprocessors and AI providers are involved, what derived data or metadata is created, what legal regimes may apply, what automated authority is granted, and whether an Australian business or practitioner can independently retain the evidence needed to understand what occurred.

  8. Examine whether changing provider–professional relationships affect employment, professional capability, competition and risk allocation as software takes over professional work; this does not mean preventing automation.

    As AI becomes embedded into those same platforms, some of the professional work previously undertaken by practitioners may increasingly be performed by the technology itself. This does not mean automation should be prevented. It does mean government should examine whether the relationship between technology providers and regulated professionals is changing in ways that affect employment, professional capability, competition and the allocation of risk.

  9. Consider how professionals will develop the judgement needed to supervise future autonomous systems if AI removes experience-building work.

    If AI progressively removes the work through which people acquire that experience, Australia needs to consider how the experienced professionals expected to supervise increasingly autonomous systems in ten or twenty years will develop their judgement.

  10. If Parliament facilitates transferring professionally significant work into software while practitioners remain accountable, consider corresponding obligations for control, evidence, auditability, consumer protection and responsibility.

    If Parliament facilitates the transfer of regulated or professionally significant work from humans into software while professional accountability remains with the practitioner, it should also consider what corresponding obligations for control, evidence, auditability, consumer protection and responsibility must move with that work.

  11. Keep adequate records and meaningful challenge routes for consequential government AI decisions; procurement must preserve government access to accountability evidence.

    Where AI contributes to consequential government decisions, there should be adequate records of what occurred and meaningful ways for affected people to question or challenge the result. Government procurement also needs to ensure that commercial technology providers cannot prevent government from obtaining the evidence needed for accountability.

  12. Distinguish information-generating AI from agentic AI authorised to interact with systems and take action.

    Parliament should distinguish between generative AI that produces information and agentic AI that is authorised to interact with other systems and take actions.

  13. Examine whether practitioners’ residual liability remains reasonably insurable as technology performs more work beyond their technical control.

    As more work is performed by technology outside the practitioner's technical control, government should consider whether the residual liability left with the practitioner remains reasonably insurable.

  14. Inform practitioners when work, corrections and behaviour improve technology; consider whether privacy, copyright and contract law adequately covers this professional knowledge, without assuming every correction is IP or training data.

    I am not suggesting that every correction is intellectual property or that a particular provider is necessarily training AI models on this information. I am saying practitioners should know when their work, corrections and professional behaviour are being used to develop or improve technology. Australia should also consider whether existing privacy, copyright and contractual laws adequately deal with this type of professional knowledge.

  15. Keep Read, Reason, Recommend, Write, Execute and Verify authorities distinct and deliberately granted, particularly for systems affecting tax, payroll, payments or authoritative financial records.

    The same principle can be expressed more precisely as: Read does not automatically mean Reason; Reason does not automatically mean Recommend; Recommend does not automatically mean Write; Write does not automatically mean Execute; and Execute should not automatically mean Verify. These are different authorities and they should be deliberately granted, particularly where an AI system can affect tax, payroll, payments or authoritative financial records.

  16. Assess cybersecurity across the complete supply chain, including external dependencies and consequences of unavailability or legal constraints, rather than just the customer-facing provider’s boundary.

    Cybersecurity assessment should also follow the complete supply chain. A platform may satisfy Australian security requirements at its own boundary while still relying on foreign cloud, AI, observability, communications, identity or other subprocessors. Each additional dependency may be well secured individually, but the combined chain can increase concentration, jurisdictional and continuity risk. For consequential financial systems, Australia should know not only where the customer-facing provider is located, but which external services are necessary for the product to continue functioning and what happens if one of those dependencies becomes unavailable or legally constrained.

  17. Ask providers about Australian infrastructure dependencies, subprocessors, cross-border flows and jurisdiction, including international data-access arrangements.

    Providers should be asked to explain their actual Australian infrastructure dependencies, subprocessors, cross- border data flows and jurisdictional exposure, including applicable international data-access arrangements such as the Australia-US CLOUD Act Agreement.

  18. Establish what leaves Australia through inference, telemetry, logs, metadata, support and subprocessors; retained/exportable activity evidence; immediately suspendable autonomy; and responses to operational services whose outputs become untrustworthy.

    The Committee should also establish what information leaves Australia through model inference, telemetry, logs, metadata, technical support and subprocessors; what evidence is retained about AI and agentic activity; whether customers can independently export that evidence; what autonomous activity can be immediately suspended; and what happens if a service remains operational but its outputs can no longer be trusted.

  19. Consider when widely used accounting, payroll, payment and financial systems become economically significant infrastructure rather than ordinary software.

    Australia should consider when widely used accounting, payroll, payment and financial systems become more than ordinary commercial software. At some point they become economically significant infrastructure.

  20. Stress-test defective mass releases, unreliable online outputs, compromised APIs, model withdrawal, cloud/identity disruption, provider failure/acquisition, price/access changes, foreign intervention and delayed-detection automated errors.

    I believe the Commonwealth should test more than a conventional cyberattack or platform outage. Scenarios should include a defective software release affecting many businesses at once; a system that remains online while producing unreliable outputs; compromise of a common integration or API; failure or withdrawal of a model relied upon by multiple services; prolonged cloud or identity-service disruption; provider failure or acquisition; substantial changes in price or access; foreign regulatory intervention; and a mass automated error that is not detected immediately.

  21. Include simultaneous record-reconstruction and verification demand in stress tests of Australia’s experienced professional and technical recovery workforce.

    The stress test should also consider human capacity. If a large number of businesses required records to be reconstructed or independently verified at the same time, would Australia still have enough experienced accountants, bookkeepers, payroll professionals, tax practitioners and technology specialists to do that work?

  22. Undertake a National AI Economic Dependency and Systemic Resilience Assessment mapping dependence on major cloud, AI, accounting, payroll, banking, payments, superannuation, identity and API infrastructure, and identifying common upstream dependencies and concentration risk.

    Undertake a National AI Economic Dependency and Systemic Resilience Assessment. Map Australia's dependence on major cloud, AI, accounting, payroll, banking, payments, superannuation, identity, API and other digital infrastructure. Identify common upstream dependencies and concentration risk, including where apparently competing Australian services rely on the same hyperscaler, model provider or infrastructure.

  23. Obtain direct evidence from both hyperscalers and the Australian software companies embedding their technology, hearing from major cloud and AI providers together with major Australian-facing accounting, payroll, financial and business software providers.

    Obtain direct evidence from both hyperscalers and the Australian software companies embedding their technology. The Committee should hear from major cloud and AI providers together with major Australian-facing accounting, payroll, financial and business software providers.

  24. Establish through that evidence the complete chain of technical dependency, data processing, subprocessors, foreign jurisdiction, delegated AI authority, auditability, liability, continuity and recovery, including what each party can independently control, stop, prove and recover.

    Evidence should establish the complete chain of technical dependency, data processing, subprocessors, foreign jurisdiction, delegated AI authority, auditability, liability, continuity and recovery, including what each party can independently control, stop, prove and recover.

  25. Align accountability with design, model selection, authority, control, evidence and prevention capacity; do not leave practitioners, businesses or consumers responsible for activity they could not reasonably see, control, verify or independently evidence.

    Make accountability follow actual capability and control. Responsibility should reflect who designed the system, selected the model, granted authority, controlled the automation, possessed the technical evidence and could prevent or stop the activity. Practitioners, businesses and consumers should not carry responsibility for automated activity they could not reasonably see, control, verify or independently evidence.

  26. Do not use historical distribution or partnership arrangements to leave professional responsibility downstream when capability and control have moved upstream to platforms.

    Historical distribution or partnership arrangements should not be used to leave professional responsibility downstream where the underlying capability and control have progressively moved upstream into the technology platform.

  27. For AI/automation affecting financial records, payroll, tax, payments, superannuation or government reporting, record relevant system/version, initiator, inputs, authority, tools/APIs, approvals, changes and actions; reject “system generated” where a precise cause is reasonably recordable.

    Establish minimum evidentiary requirements for consequential AI and automation. Where AI or automation can affect financial records, payroll, taxation, payments, superannuation or government reporting, evidence should identify, where relevant, the system or model and version, initiating actor, information relied upon, authority exercised, tools or APIs used, approvals obtained, records changed and resulting actions. A description such as "system generated" should not be sufficient where a more precise technological cause can reasonably be recorded.

  28. Require segregation of agentic authority (read, reason, recommend, write, execute, lodge or pay), separately controllable for consequential systems.

    Require segregation of agentic authority. Read should not automatically mean Reason; Reason should not automatically mean Recommend; Recommend should not automatically mean Write; Write should not automatically mean Execute; and Execute should not automatically mean Lodge or Pay. These authorities should be separately controllable for consequential systems.

  29. Require Steering, Brakes, Black Box and Backup: define authority, independently suspend autonomy, establish events and preserve/recover records for consequential autonomous systems, retaining legitimate human access when autonomous functions are disabled.

    Require Steering, Brakes, Black Box and Backup for consequential autonomous systems. Accountable organisations must be able to define authority, independently suspend autonomous activity, establish what occurred and independently preserve and recover important records. Disabling autonomous functionality should not prevent legitimate human access to underlying records.

  30. Establish independent financial-system backup/recovery standards; distinguish provider redundancy from customer backup. Enable independent preservation of transactions, source documents, attachments, audit histories and consequential automation records to reconstruct finances and the last trustworthy state.

    Establish an independent financial-system backup and recoverability standard. Provider redundancy should not automatically be treated as customer backup. Businesses and practitioners should have a practical means of independently preserving the transactions, source documents, attachments, audit histories and consequential AI or automation records needed to reconstruct their financial position and establish the last trustworthy state.

  31. Require meaningful disclosure of the complete AI and processing supply chain, so that for consequential financial systems businesses can identify material cloud providers, AI/model providers, subprocessors and jurisdictions.

    Require meaningful disclosure of the complete AI and processing supply chain. For consequential financial systems, businesses should be able to identify material cloud providers, AI/model providers, subprocessors and jurisdictions.

  32. Distinguish storage from processing and disclose material metadata, telemetry, logs and derived information; do not treat material AI-capability, authority, provider, subprocessor or data-use changes as ordinary updates.

    Disclosure should distinguish storage from processing and address material metadata, telemetry, logs and derived information. Material changes to AI capability, autonomous authority, model providers, subprocessors or data use should not be treated as ordinary software updates.

  33. Treat data sovereignty as more than data residency, considering storage, processing, controlling entities, foreign jurisdictions, subprocessors and cross-border legal access; Australian hosting alone should not establish sovereignty.

    Treat data sovereignty as more than data residency. Policy should consider where information is stored, where and by whom it is processed, which corporate entities control it, applicable foreign jurisdictions, subprocessors and cross-border legal-access regimes, including the Australia-US CLOUD Act framework. Australian hosting alone should not be treated as establishing Australian sovereignty.

  34. Recognise financial, practitioner and government-connected data as strategically significant beyond TFNs and traditional personal information.

    Recognise financial, practitioner and government-connected data as strategically significant beyond TFNs and traditional personal information.

  35. As government functions become embedded in commercial software, assess the combined data ecosystem for privacy, sovereignty, security and economic significance.

    As government functions become embedded in commercial software, the combined data ecosystem should be assessed for privacy, sovereignty, security and economic significance.

  36. Require common-dependency incident and near-miss reporting; identify correlated model, API, automation, software-release or hyperscaler failures rather than leaving them as isolated customer-support matters.

    Require systemic AI incident and near-miss reporting. Australia should be able to identify when apparently separate incidents across businesses or platforms arise from the same model, API, automation, software release, hyperscaler or other common dependency. Material correlated failures should not remain isolated customer-support matters.

  37. Stress-test online systems producing plausible errors; assess identification of affected records, stopping automation, finding the last trustworthy state, reconstructing events, restoring authoritative records and continuing essential activity.

    Stress-test integrity failure, not merely outages and cyberattacks. Testing should include systems that remain online while producing plausible but incorrect outputs, and should assess whether affected organisations can identify impacted records, stop automated activity, determine the last trustworthy state, reconstruct what occurred, restore authoritative records and continue essential activity.

  38. Where human oversight is legally or professionally required, provide reasonably necessary information, authority, evidence and controls; do not rely on “human in the loop” when technology removes visibility or control.

    Where law, regulation or professional standards require human oversight, the accountable person must have the information, authority, evidence and controls reasonably necessary to perform that oversight. "Human in the loop" should not be relied upon where technology progressively removes the human's visibility or control.

  39. Maintain equivalent protections when professional functions move into software, with no weaker standards of evidence, correction, explanation, confidentiality, dispute resolution or accountability, including for consumers using lower-cost or self-service technology.

    Maintain equivalent protections when professional functions move into software. Automation should not result in weaker standards of evidence, correction, explanation, confidentiality, dispute resolution or accountability. Consumers using lower-cost or self-service technology should not receive weaker protection because they cannot afford continuing professional supervision.

  40. Examine downstream AI liability allocation and insurability across hyperscalers, software providers, businesses, practitioners and consumers, particularly when those bearing consequences lack technical control or evidence to manage and insure risk.

    Examine contractual allocation of AI liability and insurability. Government should examine whether liability is being transferred downstream from hyperscalers to software providers, businesses, practitioners and consumers, particularly where the party carrying the consequence lacks the technical control or evidence necessary to manage and insure that risk.

  41. Have workforce policy retain sufficient accounting, tax, payroll, cybersecurity, forensic, software and financial-system expertise for independent investigation, reconstruction and recovery when automated systems fail.

    Preserve Australian human and technical recovery capability. Workforce policy should consider the expertise Australia will require when automated systems fail. Australia should retain sufficient accounting, taxation, payroll, cybersecurity, forensic, software and financial-system expertise to independently investigate, reconstruct and restore important economic records.

  42. Develop domestic assurance, RegTech, cybersecurity, forensics, provenance, backup, agent monitoring, financial continuity and domain-specific AI; sovereign capability should verify AI as well as build or consume it.

    Develop Australian capability in AI assurance and financial continuity. Support domestic capability in AI assurance, RegTech, cybersecurity, digital forensics, provenance, independent backup, agent monitoring, financial continuity and domain-specific AI. Sovereign capability should include technology capable of verifying AI, not merely building or consuming it.

  43. Require genuine exitability for economically significant digital systems; examine whether businesses and public bodies can change providers while retaining records and automated-activity evidence needed for operations and Australian legal obligations.

    Require genuine exitability for economically significant digital systems. Government should examine whether businesses and public bodies can practically change cloud, AI or software providers, including whether they can retain the records, audit evidence and historical automated activity necessary to continue operations and satisfy Australian legal obligations.

  44. Ensure people who personally carry the consequences of AI-assisted systems have direct representation in AI policy.

    People who personally carry the consequences of AI-assisted systems should have direct representation in AI policy.

  45. Make transparent the relevant commercial relationships between technology providers, industry bodies and organisations representing affected stakeholders that participate in government consultation.

    Relevant commercial relationships between technology providers, industry bodies and organisations representing affected stakeholders should be transparent where those organisations participate in government consultation.

  46. Establish an ongoing national AI systemic-risk function, with findings from AI, data centre, cybersecurity, privacy and government technology inquiries feeding continuing assessment of Australia's technological dependency chain.

    Establish an ongoing national AI systemic-risk function. Findings from AI, data centre, cybersecurity, privacy, government technology and related inquiries should feed into continuing assessment of Australia's complete technological dependency chain.

Read the original

Open in the interactive guide