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 222

Cetus AI

Submitter group: Companies.

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.

Cetus AI, an enterprise AI governance software company, argues that Australia's AI policy foundations need operational coherence and proposes five measures: a common AI evidence standard, explicit and revocable AI authority, links between governance and benefits management, procurement requiring portability and exit, and proportionate adoption pathways. It says it does not seek endorsement of a product.

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. Argues that existing government AI policies should be made interoperable at the evidence level rather than replaced.

    The policy opportunity is therefore not to replace existing frameworks, but to make them interoperable at the evidence level.

  2. Argues that accountability requirements should scale with the consequence and reversibility of an AI action rather than the label on the technology.

    These requirements should be implemented proportionately. A low-consequence internal drafting assistant should not carry the same burden as a system affecting eligibility, safety, payments, enforcement, employment or access to essential services. The decisive factor should be the consequence and reversibility of the action, not the marketing label attached to the technology.

  3. Argues that public trust depends on AI being safe and lawful and on public resources producing demonstrable value.

    Public trust depends not only on whether AI is safe and lawful, but also on whether public resources are producing demonstrable value.

  4. Argues that data sovereignty should be treated as a multidimensional control objective rather than a binary hosting label.

    Data sovereignty should be treated as a multidimensional control objective rather than a binary hosting label.

  5. Argues that operational accountability should be made achievable for organisations that cannot build a large governance function.

    The National AI Centre already provides practical guidance and support.[7] The next step should be to make operational accountability achievable for organisations that cannot build a large governance function.

    A program corrected obvious text-recognition errors in this quote for display.

Recommendations

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

  1. Retain workload, payload, identity, enforcement and consequential-decision responsibility in customer/agency systems; technology may present evidence but should not manufacture causation, issue accounting/legal opinions, certify compliance or replace authorised decision-makers.

    Customer and agency systems should retain responsibility for workloads, payloads, identity, enforcement and consequential decisions. Technology can preserve and present evidence; it should not manufacture causation, issue an accounting or legal opinion, certify compliance, or replace an authorised decision-maker.

  2. Establish a vendor-neutral minimum evidence standard for material AI use cases, extending existing registers, impact assessments, technical standards, procurement and benefits-management requirements.

    The Australian Government should establish a vendor-neutral minimum evidence standard for material AI use cases. The standard should extend—not duplicate—the current use-case register, impact-assessment, technical-standard, procurement and benefits-management requirements.

    A program corrected obvious text-recognition errors in this quote for display.

  3. Define minimum evidence fields and lifecycle events without requiring sensitive-data centralisation or a single platform; allow evidence retention in appropriate systems connected through common identifiers and export formats.

    The standard should define the minimum evidence fields and lifecycle events that an agency must be able to produce. It should not require agencies to centralise sensitive operational data or adopt a single technology platform. Agencies should be able to retain evidence in appropriate systems while using common identifiers and export formats to connect the record.

  4. Include use-case identity, purpose, owner, affected users, materiality and lifecycle; authority, permissions, approvals, constraints, expiry and revocation; and data classifications, models, providers, tools, interfaces, locations, dependencies and versions.

    A minimum standard should address: - Use-case identifier, business purpose, accountable owner, users Identity and purpose or populations affected, materiality and lifecycle state Authori Approved scope, permissions, delegations, human-approval arty) requirements, constraints, expiry and revocation conditions tect Relevant data classifications, models, providers, tools, interfaces, Data and nology locations, dependencies and material version changes

  5. Include assessments, policies, controls, exceptions, reviews, incidents, remediation, monitoring, drift, human interventions, service changes, decommissioning, baselines, targets, costs, outcomes, assumptions, dependencies, validation and authorised portfolio decisions.

    Applicable impact assessment, policies, control decisions, Gov ce exceptions, reviews, incidents and remediation Operations Monitoring results, material drift, human interventions, service changes and decommissioning evidence Approved baseline, targets, cost sources, observed outcomes, Value assumptions, dependencies, Finance or business validation and authorised portfolio decisions

  6. Link rather than necessarily relocate records across risk, procurement, architecture, service management, Finance and business systems using common identifiers, respecting security classifications, privacy, privilege and operational separation.

    The standard should use common identifiers that allow an agency to link—not necessarily relocate—records held across risk, procurement, architecture, service management, Finance and business systems. This would support internal oversight, external review and learning across agencies while respecting security classifications, privacy, legal privilege and operational separation.

  7. Have DTA lead an open specification with proportionate conformance guidance and reference implementations, consulting Finance, the AI Safety Institute, security/privacy authorities, agencies, states/territories, Australian industry, researchers, unions and civil society.

    The DTA should lead the standard in consultation with the Department of Finance, the Australian AI Safety Institute, security and privacy authorities, agencies, state and territory counterparts, Australian industry, researchers, unions and civil society. The result should be published as an open specification with proportionate conformance guidance and reference implementations.

    A program corrected obvious text-recognition errors in this quote for display.

  8. For consequential uses, require named ownership, machine/service identity where applicable, least-privilege access, documented actions, human review/escalation/suspension/re-authorisation conditions, revocable access, service removal, change monitoring and action-linked evidence of valid authority.

    For consequential use cases, agencies should therefore be able to demonstrate: a named organisational owner and, where applicable, a distinct machine or service identity; least-privilege access and a documented scope of permitted actions; conditions that require human review, escalation, suspension or re-authorisation; an effective ability to restrict or revoke access and remove the system from service; monitoring that can identify material changes in behaviour, context, dependencies or risk; and an action-linked record showing that the relevant authority and conditions were valid when a consequential action was attempted.

  9. Scale requirements to consequences and reversibility: low-consequence internal drafting should carry less burden than systems affecting eligibility, safety, payments, enforcement, employment or essential-service access.

    These requirements should be implemented proportionately. A low-consequence internal drafting assistant should not carry the same burden as a system affecting eligibility, safety, payments, enforcement, employment or access to essential services. The decisive factor should be the consequence and reversibility of the action, not the marketing label attached to the technology.

  10. Keep enforcement and intervention responsibility with agency systems and authorised officials; prevent AI from being the final arbiter of its own authority.

    The AI system should not be the final arbiter of its own authority. Agency systems and authorised officials should retain enforcement and intervention responsibility.

  11. For material government AI investments, trace problems/alignment, approved baselines/targets/benefits, Senior Responsible Official/Benefit Owners, approved costs/assumptions/dependencies/comparison periods, results/governance events, attribution/uncertainty, Finance/business validation through existing authorised processes, and Scale/Optimise/Review/Retire decisions with owners/review dates.

    For material Government AI investments, the Committee should recommend a traceable value-accountability chain containing: 1. the public or operational problem and its strategic alignment; 2. the approved baseline, target measures and expected benefits; 3. the accountable Senior Responsible Official and Benefit Owners; 4. approved cost sources, assumptions, dependencies and comparison periods; 5. observed operational results and material governance events; 6. a documented attribution method and uncertainty or limitation statements; 7. Finance and business validation through existing authorised processes; and 8. a recorded decision to Scale, Optimise, Review or Retire, with an owner and next review date.

  12. Prevent automatic software declarations of ROI/causation and treatment of forecasts/estimates as realised benefits; preserve Finance/business-owner authority over attribution/benefit recognition and decision-forum authority over investments.

    This chain should not enable software to declare ROI or causation automatically. Forecasts and operational estimates are not realised benefits. Finance and business owners must retain authority for attribution and benefit recognition, and decision forums must retain authority over investment choices.

  13. Start pilots with one workflow, approved baseline, cost source and Benefit Owner and a defined review period; validate the evidence-and-decision method before broader portfolio adoption.

    A practical starting point is intentionally narrow: one workflow, one approved baseline, one cost source, one Benefit Owner and a defined review period. The purpose of a pilot should be to validate the evidence-and-decision method before considering broader portfolio adoption.

  14. Reinforce AI procurement guidance with model procurement clauses covering operational evidence and exit.

    Those expectations should be reinforced through model procurement clauses that address operational evidence and exit from the outset.

  15. For material AI procurements, require portable evidence, documented dependencies, agency-controlled retention/deletion/access/disclosure, material-change notices, and ways to apply and evidence agency controls, exceptions, human approvals and stop conditions.

    For material AI procurements, agencies should require suppliers to support: Intended public outcome Exportable, machine-readable records sufficient to support Evidence portability agency oversight, transition and review . Clear identification of models, tools, data pathways, Documented dependencies subprocessors and material external services cus co) led retention Agency control over retention, deletion, access and lawful disclosure of governance evidence Ch visibility Notice ea for material model, service, policy, location or Practical ways for agency controls, exceptions, human approvals and stop conditions to be applied and evidenced

    A program corrected obvious text-recognition errors in this quote for display.

  16. Require tested pathways to remove access, preserve required records, migrate services and avoid unmanaged residual authority; authorise and test data flows, controls and evidence mappings in the specific agency environment.

    Tested pathways to remove access, preserve required records, Policy and decision interfaces Exit and continuity arrangements migrate services and avoid unmanaged residual authority Confirmation that data flows, controls and evidence mappings Environment validation have been authorised and tested in the specific agency environment

  17. Treat data sovereignty as multidimensional control: assess access, processing and retention locations, law and contracts, identity and encryption control, foreign dependencies, record exportability and safe continuation or exit.

    Data sovereignty should be treated as a multidimensional control objective rather than a binary hosting label. Relevant considerations include who can access data and evidence, where they are processed and retained, which law and contract terms apply, who controls identity and encryption keys, which foreign services or support arrangements are dependencies, how records can be exported, and whether the agency can continue or exit safely.

  18. Avoid turning sovereign capability into technological isolation in procurement, so Australian firms can provide specialised components within broader architectures.

    Procurement should also avoid turning sovereign capability into technological isolation. Australian firms should be able to provide specialised governance, data, security, evaluation, integration and value-accountability components within broader architectures.

  19. Publish proportionate reference patterns, model clauses and bounded-pilot templates distinguishing productivity tools, business-process AI, citizen-facing systems, regulated decisions and tool-using agents; scale evidence burdens with consequence, scale, sensitivity and reversibility.

    The Australian Government should publish a set of proportionate reference patterns, model clauses and bounded- pilot templates. These should distinguish between low-consequence productivity tools, business-process AI, citizen- facing systems, regulated decisions and agentic systems with tool access. The evidence burden should increase with consequence, scale, sensitivity and reversibility.

    A program corrected obvious text-recognition errors in this quote for display.

  20. Consider providing small-organisation use-case and owner records, authority/oversight templates, data/model/supplier records, change and re-authorisation triggers, baselines and benefits profiles, procurement clauses for portability and exit, and post-pilot decision records.

    A practical package could include: © aminimum use-case record and owner declaration; e asimple authority and human-oversight template; e a data, model and supplier-dependency record; © a material-change and re-authorisation trigger list; « a baseline and benefits profile suitable for small organisations; © model procurement clauses for evidence portability and exit; and e a short post-pilot decision record.

  21. Co-design reference implementations through time-bounded pilots with agencies, Australian providers, advisers, researchers, SMEs and regional organisations; publish lessons where confidentiality/security allow; test existing tools and architectures before defaulting to a central platform.

    Reference implementations should be co-designed through time-bounded pilots involving agencies, Australian technology providers, advisory firms, researchers, SMEs and regional organisations. Lessons should be published where confidentiality and security allow. Government should test the standard against existing tools and architectures rather than creating a new central platform by default.

  22. Take a staged approach that builds on current policy rather than introducing a parallel regime.

    Cetus AI recommends a staged approach that builds on current policy rather than introducing a parallel regime.

    A program corrected obvious text-recognition errors in this quote for display.

  23. Have DTA convene a cross-functional group to define materiality, evidence domains and common identifiers without centralising sensitive data; map to current frameworks, identify duplication and preserve statutory and agency authority.

    DTA convenes a cross-functional working - a 1. Define group to define materiality, evidence No requirement to centralise sensitive data domains and common identifiers Map the evidence standard to current AI 2. Map policy, impact assessment, technical Identify duplication and explicitly preserve standard, procurement and Benefits existing statutory and agency authority

    A program corrected obvious text-recognition errors in this quote for display.

  24. Pilot with stop conditions and review; evaluate administrative burden, evidence quality, interoperability, security, decision usefulness and SME proportionality.

    3. Pilot facing service support and a bounded stop cond and ined review agentic workflow Assess administrative burden, evidence - —4. Evaluate quality, interoperability, security, decision owe Su and publication Gi usefulness and SME proportionality

  25. Publish the standard, model clauses, reference patterns and transition guidance; phase adoption by materiality and agency readiness; regularly review as technology, law and experience evolve, with transparent change control and stakeholder consultation.

    5. Implement Publish the standard, model clauses, Phased adoption based on materiality and . reference pattems and transition guidance —_ agency readiness Review the standard regularly as s - Transparent change control and 6. Improve technology, law and operating experience stakeholder consultation develop

Read the original

In some quotes a program corrected obvious text-recognition errors for display.

Open in the interactive guide