Guidance for AI Adoption: Implementation practices

There are 2 versions of the guidance. This version, Implementation practices, is for: 

• organisations who are developing AI models or systems, or using them in multiple ways
• organisations operating in high-risk or sensitive environments
• governance professionals
• professionals with technical AI expertise
• professionals looking to align with best-practice AI governance. 

Organisations just starting out in AI adoption and professionals who are new to AI and AI governance should use the use the other version of the guidance, Foundations. 

How to use this guidance

This guidance applies to both developers and deployers of AI systems. Where practices are particularly relevant to either developers or deployers, this is marked with a corresponding (DEP) or (DEV).

Find all the definitions for this guidance in the terms and definitions.

1. Decide who is accountable 

AI systems can make automated decisions that significantly impact people, communities and businesses. Overall, your organisation is ultimately accountable for how and where AI is used, AI complexity can create gaps where no one takes clear responsibility for outcomes. 

Accountability is the first step to using AI responsibly. 

1.1 Accountable people

2. Understand impacts and plan accordingly

Because AI systems can operate at speed and scale, their potential impacts are often magnified. Without careful planning, a single AI system can lead to widespread negative outcomes, such as unfair decisions or the provision of inaccurate information. 

For example, AI systems can learn from and amplify existing issues such as unwanted bias in data. This can lead to unfair decisions or inappropriate generated content that could affect many people. If an AI system used for shortlisting in hiring has a bias problem, it could unfairly reject hundreds of qualified candidates before anyone notices.

To use AI responsibly, organisations need to understand, plan for and monitor potential impacts of AI systems. Those affected should be able to raise complaints and get help.

2.1 Identify and engage stakeholders

2.1.1 Identify and document key types of stakeholders (such as employees and end users) that may be impacted by the organisation’s development and deployment of AI, and their needs.

2.1.2 Prioritise, select and document which stakeholder needs will be addressed in organisational policies and procedures.

2.1.3 Document and communicate the organisation’s commitment to preventing harms to people from AI models and systems and upholding diversity, inclusion and fairness.

2.1.4 Document the scope for each AI system, including intended use cases, foreseeable misuse, capabilities, limitations and expected context.

2.1.5 For each AI system, engage stakeholders to identify and document the potential benefits and harms to different types of stakeholders, including:

• impacts to vulnerable groups
• risks of unwanted bias or discriminatory outputs
• use of an individual’s personal information
• where the system makes or influences a decision about a person or group of people.

2.1.6 For every documented risk of harm to affected stakeholders, conduct appropriate stakeholder impact analysis.

2.1.7 Monitor for potential harms by engaging affected stakeholders for each AI system on an ongoing basis to identify new stakeholders, including end users throughout the AI lifecycle.

2.1.8 Create processes to support ongoing engagement with stakeholders about their experience of AI systems. Identify vulnerable groups and support appropriately. Equip stakeholders with the skills and tools necessary to give meaningful feedback.

2.2 Establish feedback and redress processes

2.2.1 Create, document and communicate a process for:

• Potentially affected stakeholders to raise concerns, challenges, or requests for remediation and receive responses (for example, a human rights grievance and remediation mechanism). This includes when and how frequently to communicate, the level of detail to provide and the communication needs of stakeholders, considering the level of AI knowledge and any regulatory requirements.
• Evaluation of contestability requirements of both internal and external stakeholders and interested parties including accessibility needs.

2.2.2 Implement and document system‑level mechanisms to enable contestability of AI use and decisions, enabling stakeholders to understand, challenge and appeal AI use and decisions. These mechanisms must be accessible, understandable and available to users at the appropriate time during interaction with an AI system. Consider mechanisms to share information regarding end user contests and any redress with deployers of AI systems and models (DEV).

2.2.3 Implement and document mechanisms to enable deployers to escalate feedback, report unexpected behaviours, performance concerns or realised harms and support improvements to models or systems (DEV).

2.2.4 Ensure people who monitor and review affected stakeholder feedback can trigger recourse and redress processes where there is an obligation to do so.

2.3 Monitor for systemic issues

2.3.1 Monitor and evaluate contestability and redress processes to identify and address systemic risks as well as improve effectiveness of these processes.

2.3.2 Create, document and communicate a process to review and evaluate stakeholder contests of AI system use across the organisation including any concerns raised by affected stakeholders and requests for information. 

3. Measure and manage risks: implement AI-specific risk management

AI risks fundamentally change depending on the type and complexity of your AI systems. Risks often emerge from how the AI system behaves in different situations and use-cases, rather than only from software updates. They can rapidly amplify smaller issues into significant problems.

For example, an AI chatbot that answers simple questions during business hours, when it can be monitored by a staff member, is a low-risk use of AI. The risks expand, however, if that chatbot operates 24/7, without human oversight, and answers more complex questions. 

To use AI responsibly, organisations need to be able to identify and manage its risks. 

3.1 Establish a fit-for-purpose risk management framework

3.1.1 Create and document:

• a risk management framework that addresses the specific characteristics and risks of AI systems
• organisational‑level risk tolerance and criteria to determine acceptable / unacceptable risks for the development and deployment of AI systems. This should include the significance and likelihood of potential harms to affected stakeholders in line with the AI policy and objectives.
• AI impact assessment, risk assessment and risk treatment processes, including criteria for reassessment over the lifecycle of an AI system. Identify and document any specific use cases or qualities of AI systems that represent an unacceptable risk to stakeholders or the organisation, in line with the organisation’s risk tolerance.

3.1.2 Ensure that risk management processes include steps to identify, assess and treat risks arising from other parties in the AI supply chain, such as third‑party developers and third party deployers. Specific risks relating to open‑source AI models, systems and components should be considered by both providers and consumers of these technologies.

3.1.3 Adopt or develop clear and consistent reporting formats, such as data sheets, model cards, or system cards, to communicate appropriate risk management outcomes, including residual risks, to relevant stakeholders (DEV).

3.2 Assess AI system risks

3.2.1 Establish a triage system to determine which AI systems may pose an enhanced or unacceptable risk, aligned to the organisation’s context and risk tolerance (see Foundations Triage template).

3.2.2 Perform and document a risk assessment and evaluation for the specific requirements, characteristics and documented use cases of each AI system, including systems developed or procured from third party suppliers. 

3.2.3 In undertaking AI system risk assessments, take the following steps to evaluate the likelihood and consequence of each risk as well as the consequence of not deploying the AI system: 

• Identify potential severity and likelihood of harms to stakeholders, drawing on the Stakeholder Impact Assessment (see 2.1.1 – 2.1.6).
• Identify legal, commercial and reputational risks such as failing to meet legal obligations, organisational commitments to ESG, diversity, inclusion and accessibility or programs supporting diversity, equity and fairness.
• Consider the potential amplified and emerging data governance risks across each phase of the AI system lifecycle including before and after model training.
• Analyse risks systemically using risk models to identify the sources and pathways through which AI systems could produce the identified risks.
• Compare the estimated value or level of identified risks to pre‑determined organisational risk criteria (see 3.1.1) or those defined by regulatory bodies or stakeholders.
• Document any specific use cases or qualities that represent an unacceptable level of risk to stakeholders or the organisation.
• Communicate risk assessments in clear reporting formats to relevant stakeholders.

3.3 Implement controls for AI system risks

3.3.2 Communicate risk treatment plans in clear reporting formats to relevant stakeholders.

3.3.3 Create and document a deployment plan which includes the response, recovery and communications for the realization of residual risks.

3.3.4 Research, document and implement leading practices in safety measures as safeguards, as appropriate for identified risks (DEV).

3.4 Monitor and report incidents

3.4.1 Track, document and report relevant information about serious incidents and possible corrective measures to relevant regulators and/or the public in a reasonable timeframe. Reporting near‑misses and corrective measures is good practice. Communication of corrective measures should consider privacy and cybersecurity risks.

3.4.2 Create and document a process to evaluate and fulfil reporting and disclosure obligations such as those under the Online Safety Act relevant to AI systems usage, including documentation of safety measures implemented such as notices and incident reporting.

3.4.3 Conform to and document data breach reporting requirements and liabilities from related standards. For example, under the Notifiable Data Breaches scheme of the Office of the Australian Information Commissioner.

3.4.4 Maintain two‑way communication between developers and deployers for incident reporting, sharing performance insights and coordinating responses to identified issues.

3.4.5 Monitor and evaluate risk assessments and treatment plans on a regular, periodic basis or when a significant change to the use case or the system occurs, or new risks are identified. This includes responding to impact assessments or insufficient risk treatment plans.

3.4.6 Monitor and evaluate the overall effectiveness of risk management processes and continually improve them. 

4. Share essential information

People should know when they’re interacting with AI and understand when AI decisions affect them. For example, when a customer is receiving services and guidance from a chatbot, they should know this is not a human specialist. 

To use AI responsibly, organisations need to tell users and stakeholders when and how they’re interacting with AI. 

4.1 Maintain an AI register

4.1.1 Create and maintain an up‑to‑date, organisation‑wide inventory of each AI model and system, with sufficient detail to inform key stakeholders and support future conformance assessments, including:

• accountable people
• purpose and business goals
• capabilities and limitations of the AI model and/or system
• origin, fine‑tuning and updates where applicable
• technical requirements and components
• datasets and their provenance used for training and testing
• acceptance criteria and test results
• any impact and risk assessments and outcomes
• identified risks, potential impacts and the risk treatment plan
• any system audit requirements and outcomes
• dates of review.

See the AI register template for additional guidance.

4.2 AI system transparency and explainability

4.2.1 Create, document, implement and communicate a policy and process for how to transparently communicate:

• to AI users, and affected stakeholders that engage directly with AI systems, AI‑enabled decisions or AI‑generated content about when and how they will be informed about AI development, deployment and use in the organisation.
• the capabilities, limitations and potential risks of the AI systems that users and affected stakeholders may engage with. This should include when and how frequently to communicate, the level of detail and the level of AI knowledge of AI users and affected stakeholders. It should also address communication obligations and the accessibility needs of AI users and affected stakeholders and incorporate feedback mechanisms where appropriate.

4.2.2 For each AI system, evaluate and document:

• the transparency requirements for each user and affected stakeholder group (see 2.1.1 and 2.2.1)
• the transparency and explainability system requirements and measures – including for third‑party ‑provided systems – dependent on use case, stakeholder requirements and risks arising from disclosure.
• how accessibility obligations and commitments are met by implementing human‑centered design.

4.2.3 Create, document and implement organisational processes and transparency mechanisms proportionate to the risks arising from the diverse, evolving, and alternative uses of GPAI beyond predefined applications, including their potential for unexpected and hard‑to‑explain behaviours. (GPAI DEV/DEP).

4.2.4 Wherever possible choose more interpretable and explainable systems.

4.2.5 Wherever possible, provide reasonably interpretable and explainable AI systems and models to accountable people within the organisation or downstream deployers to enable them to meet their own regulatory obligations (DEV).

4.2.6 Conduct internal testing of the AI model and/or system’s capabilities and limitations. Clearly communicate results to deployers prior to deployment. (DEV).

4.3 Supply chain transparency

4.3.1 Document or request from upstream providers the technical details of the system or model that may be required to meet the needs of users within the organisation or stakeholders.

4.3.2 Share as much of the following information as possible about AI models and systems with downstream deployers (while protecting commercially sensitive information and meeting legal compliance) (DEV):

• Technical details such as model architecture, description of data, components and their characteristics
• Test methods, use cases and testing resulting
• Known limitations, risks and mitigations (such as potential bias and corrective actions) and external audit findings
• Data management processes for training and testing data including data quality, meta data, and provenance
• Privacy and cybersecurity practices including conformance to standards and best practice
• Transparency mechanisms implemented for AI‑generated content, interactions and decisions
• Document and share the following key information for GPAI systems with downstream organisations, stakeholders, researchers and regulators (GPAI DEV):
  • Training data sources and compliance details with relevant privacy, intellectual property and copyright laws
  • Model cards and system cards, including risk assessment results particularly evaluation of dangerous and emerging capabilities in the deployment and scaffolding context of tool access and agent design.
• Restricted and managed access to model weights and other associated artefacts.

4.3.3 Share as much of the following information as possible about AI systems with upstream developers (while protecting commercially sensitive information and meeting privacy obligations) (DEP) (See also incident reporting 3.4.1 and 3.4.4).

• Issues, faults, failures, incidents and any other observed risks that can be addressed by developers
• Any unexpected and unwanted bias resulting from use of the system.
• Ensure you’ve included the required information in contracts with suppliers of systems, including when to update information.

4.4 AI-generated content transparency

4.4.1 Implement fit‑for‑purpose and proportionate transparency mechanisms for AI generated content as set out in the Being clear about AI‑generated content guidance (forthcoming).

5. Test and monitor

AI systems can change their behaviour over time or act in ways that are less predictable than conventional software. For example, an AI system that worked well last month might start giving different answers today if it is trained on additional data. 

To use AI safely, organisations should test and monitor their AI systems. 

5.1 Pre-deployment testing

5.1.1 Establish oversight mechanisms to review and approve testing methodologies and results, and to monitor system performance, user feedback and operational impacts post‑deployment.

5.1.2 Define and communicate clear acceptance criteria and test methodologies that reflect the intended use, context and potential risks (as identified in Essential Practice 3). Conduct predeployment testing.

5.1.3 Clearly document tests and outcomes to support external audits and oversight

5.1.4 When testing is conducted by upstream providers during the development of the AI system and model, request test methodologies and results from the provider and ensure its alignment with your acceptance criteria.

5.1.5 Obtain documented deployment authorisation and rationale from the accountable person for the AI system based on test results.

5.2 Monitor system performance

5.2.1 Establish monitoring systems for each AI system to track key performance metrics and indicators relevant to the identified risks.

5.2.2 Implement a deployment process for AI systems that maps business targets to system performance metrics for both internal and third‑party developed systems.

5.2.3 Establish and document response processes for addressing all foreseeable issues and harms during system operation.

5.2.4 Establish regular system performance review cycles with stakeholders and subject matter experts to evaluate testing criteria, effectiveness and outcomes.

5.2.5 For each AI system, create and document monitoring requirements including human oversight prior to deployment and evaluate as part of continuous improvement cycle.

5.3 Conduct additional testing proportionate to risk

5.3.1 Conduct safety evaluations that scale with model capabilities (GPAI DEV), for example: assessment for cyber‑offensive capabilities and vulnerabilities; testing for potential chemical, biological, radiological and nuclear information risks; evaluation of model behaviours beyond intended use cases; testing for jailbreaking or prompt manipulation; data privacy risks; or comprehensive red teaming to identify vulnerabilities.

5.3.2 For use cases requiring enhanced practices and GPAI systems, conduct independent (internal or external) and thorough review of testing and evaluation methodologies and results. Document and report any issues to the accountable person for the system.

5.3.3 Create a process for and determine whether an AI system requires regular auditing, appropriate to the level of risk identified by its risk assessment. Conduct audits when required.

5.4 Implement robust data and cybersecurity measures

5.4.1 Implement and evaluate the effectiveness of policies and procedures in addressing AI‑specific risks and adapt as necessary:

• Data governance processes covering the use of data with AI models and systems. This includes the management of data usage rights for AI including intellectual property (including copyright), Indigenous Data Sovereignty, privacy, confidentiality and contractual rights.
• Privacy policies covering the collection, use and disclosure of personal or sensitive information by AI models and systems, including for model training purposes. This needs to support teams to comply with the Australian Privacy Principles for all AI systems.
• Cybersecurity processes to cover the emerging and amplified risks of AI systems interaction with existing systems and data, such as AI systems unintentionally exposing sensitive information or bypassing security controls. This includes application of the Essential Eight Maturity Model for cybersecurity risks to AI systems.

5.4.2 For each AI use case:

• Define and document the data quality, data/model provenance and data preparation requirements.
• Understand and document the data sources and collection processes each AI model / system relies on to function including personal and sensitive data. Put in place systems to manage the data and document the data used to train and test each AI model or systems and data used for inference.
• Define and document processes for protecting AI models and systems to address emerging cybersecurity and privacy risks (DEV).
• Where appropriate, report to relevant stakeholders on data, model and system provenance.
• Document how the Australian Privacy Principles have been applied including in models and systems developed by third parties.
• Document data usage rights including intellectual property (including copyright), indigenous data sovereignty privacy confidentiality and contractual rights.
• Monitor for and detect any leakage of personal and sensitive information from AI models and systems.

6. Maintain human control

Unlike traditional software that follows explicit instructions, AI systems learn patterns from data and make their own opaque decision logic. This means they need human oversight to make sure they operate safely. For example, while regular software does exactly what you program it to do, AI might interpret your instructions differently than you intended. 

To responsibly use AI, organisations need to make sure a human appropriately oversees any AI systems in use. The person overseeing your AI systems should know how to do so appropriately, and what they need to do to override the system if something goes wrong. 

6.1 Maintain human oversight and control

6.1.1 Maintain operational accountability, capability and human oversight throughout the lifecycle of AI systems.

6.1.2 Implement mechanisms to enable human control and intervention during the operation of the AI system (DEV).

6.1.3 Implement mechanisms to enable human oversight and intervention to address systemic risks and emerging capabilities such as capability evaluation, training, pause, independent oversight, dynamic guardrails and tool/system access controls (DEV).

6.1.4 Ensure appropriate training is provided to anyone overseeing or using AI systems to understand each system’s capabilities, limitations and failure modes and when human intervention is needed.

6.2 Decommission when appropriate

6.2.1 Define and determine:

• the criteria or reasons that termination of an AI model or system might need to occur, and at what point intervention should take place
• the most appropriate role or person to oversee the intervention and decommissioning process
• whether the model or system is essential to any critical infrastructure or service delivery
  • Assess the risks and impacts of shutting down the AI model or system, including impacts on end‑users and interdependencies with other integrated systems on which the model, data or outputs rely to function.
  • Develop a timeframe and treatment plan to minimise impacts or disruption caused by the decommissioning process.
  • Determine a method to extract data and for the return or deletion of assets. Establish which information should be preserved for record keeping purposes.

6.2.2 Create a process for how your organisation will inform relevant parties (such as  employees, customers, and upstream or downstream parties) within a reasonable timeframe of the retirement or shutdown of an AI model or system. Establish a channel for people to raise concerns, request support and receive responses.

6.2.3 Determine whether alternative systems or processes will need to be provided to address any issues or gaps.

6.2.4 Maintain alternative pathways for critical functions so operations can continue if AI systems malfunction and/or are taken offline.