Glossary

A

A2A (Agent2Agent)

An open protocol that lets one organisation’s AI agents hand tasks to another’s.

Accountable official

In government, the official accountable for implementing Policy v2.0 in the agency, with the DTA told of changes.

Accountable owner

The single named person who answers for a use case’s outcomes and makes sure its controls run. In government, the ‘accountable use-case owner’ under Policy v2.0.

ACM (Association for Computing Machinery)

A professional computing society; publisher of the journal in which CSIRO’s Responsible AI Pattern Catalogue appeared.

ADM (automated decision-making)

A computer program making, or doing something substantially and directly related to making, a decision. Under the Privacy Act from 10 December 2026, certain automated decisions must be explained in privacy policies.

AGD (Attorney-General’s Department)

Co-author of the Automated Decision-Making Better Practice Guide.

Agent

An AI system that can take actions through tools (send, file, update records, run code), not only produce text. ‘Agent’ is the word Renly uses for its AI.

AI (artificial intelligence)

Software that produces predictions, recommendations, content or decisions from data in ways that are not fully specified by explicit rules. Includes generative assistants, machine learning models and agents.

AI centre of excellence

A small enabling team that runs the register and lifecycle, supports owners and maintains templates. It enables; it does not own use cases.

AI governance committee

The cross-functional body with decision rights over AI use cases, tiers, gates, approved tools and exceptions.

AI incident

A harmful, unlawful or materially wrong AI output acted on, a privacy or security breach involving AI, sustained behaviour outside tolerance, or consequential AI found operating outside the register.

AI RMF

NIST’s AI Risk Management Framework. Its four functions are written in capitals: GOVERN, MAP, MEASURE and MANAGE.

AI use case

A specific use of AI for a purpose in a business context, such as ‘drafting replies to customer emails with an assistant’. The unit that is registered, owned, tiered and assessed.

AICD (Australian Institute of Company Directors)

The professional body for directors; co-publisher of a director’s guide to AI governance.

ANAO (Australian National Audit Office)

The Commonwealth’s independent auditor, which now audits AI governance.

API (application programming interface)

A documented way for one software system to request data or actions from another, used to connect systems and automate exports without a person clicking through screens.

APP (Australian Privacy Principle)

One of the 13 principles in the Privacy Act that govern handling of personal information. ‘APP entity’ means an organisation or agency bound by them.

APRA (Australian Prudential Regulation Authority)

The regulator of banks, insurers and superannuation funds. Its CPS (cross-industry prudential standard) documents include CPS 230 and CPS 234.

APS (Australian Public Service)

The Commonwealth public service.

ASD (Australian Signals Directorate)

The agency responsible for the Information Security Manual and cyber security guidance.

ASIC (Australian Securities and Investments Commission)

The corporate and financial services regulator.

ATO (Australian Taxation Office)

The Commonwealth tax agency; the subject of an ANAO audit of AI governance.

Autonomy

How much a system may do without a person’s approval. Governing Intelligence describes five rungs: assistive, augmentative, supervised, autonomous and orchestrated.

B

BYOK (bring your own key)

Encrypting stored data or credentials with a key the customer controls.

BYOLLM (bring your own large language model)

Routing model calls to a model provider the customer contracts directly.

C

CAIO (Chief AI Officer)

The senior executive who leads AI adoption and governance. The APS AI Plan expects every Australian Government agency to appoint one in 2026.

Calibration

Checking that assessors apply the rating rules the same way, using worked examples.

Champion

A staff member in a business area, with time allocated, who helps colleagues use AI well and feeds back problems.

CIO (Chief Information Officer)

The executive responsible for information technology.

CISO (Chief Information Security Officer)

The executive responsible for information and cyber security.

CMMI (Capability Maturity Model Integration)

A widely used process maturity model whose five-level structure informed the levels in this model.

Contestability

The ability of a person affected by a decision to challenge it and have a human review it.

CPS (cross-industry prudential standard)

APRA’s binding standards, such as CPS 230 Operational Risk Management.

Criterion

One of 40 specific capabilities scored in the assessment.

CSIRO (Commonwealth Scientific and Industrial Research Organisation)

Australia’s national science agency; its Data61 unit publishes the Responsible AI Pattern Catalogue.

CSV (comma-separated values)

A plain spreadsheet file format.

CTO (Chief Technology Officer)

The executive responsible for technology strategy and engineering.

D

Data loss prevention

Technical controls that detect or block sensitive information leaving approved systems.

Decision record

The record that travels with an AI-assisted decision: subject, policy applied, model and version, sources, confidence, human approvals and rationale. Governing Intelligence calls it a decision passport.

Descriptor

The description of what a criterion or domain looks like at a given level.

Discovery sweep

A deliberate search for AI already in use, including features inside purchased software and staff use of public tools, run under an amnesty.

Domain

One of eight areas of AI capability assessed by the model, each with five criteria.

DTA (Digital Transformation Agency)

The Australian Government agency that owns the Policy for the responsible use of AI in government.

E

Edition

The version of the reference set applied: Australian Government or organisation. The criteria are the same; references and obligations differ.

Embedded AI

AI features inside software the organisation already licenses, often switched on by the vendor.

EU (European Union)

Relevant where an organisation provides AI on the EU market or its AI’s output is used in the EU.

Evaluation suite

A set of test cases with agreed pass criteria that is run before release and after every change.

Evidence confidence

Evidenced (seen), asserted (stated but not seen) or not known. Asserted scores are capped at Level 2.

Evidence ID

The reference number given to each item in the evidence log. Every score at Level 3 or above cites at least one.

Evidence Passport

Renly’s sealed evidence record for each response its agents give.

F

FOI (freedom of information)

The right of access to government documents under the Freedom of Information Act 1982.

G

Gap

Target minus current score.

Gate

A go or no-go decision point in the lifecycle, decided on evidence by the right authority and recorded.

GenAI (generative AI)

AI that generates text, images, code or other content.

GOVERN, MAP, MEASURE, MANAGE

The four functions of the NIST AI Risk Management Framework, cited with their category numbers (for example GOVERN 1).

Governing Intelligence

Renly’s AI governance framework for Australian Government, published at renly.ai/governing-intelligence, with a free governance assessment.

H

Horizon

A roadmap period: 0 to 90 days, 3 to 6 months, 6 to 12 months, 12 to 24 months.

HR (human resources)

The people and capability function.

Human participation point

A designed point where a named person reviews, approves or decides, with defined information and authority.

I

ICT (information and communications technology)

The government term for IT, including communications systems.

Impact assessment

A structured assessment, before deployment, of an AI use case’s effects on people, privacy, fairness, security and lawfulness.

Initiative

A defined piece of work in the library that lifts one or more criteria from one level to another.

IP Australia

The Commonwealth agency that administers intellectual property rights; the subject of an ANAO audit of AI use.

IRAP (Infosec Registered Assessors Program)

ASD’s program of registered assessors who assess systems and cloud services against the ISM.

ISM (Information Security Manual)

ASD’s cyber security controls for Australian Government systems, including AI controls.

ISO/IEC

The International Organization for Standardization and the International Electrotechnical Commission, which publish joint AI standards such as ISO/IEC 42001.

IT (information technology)

Computer systems, software and networks, and the teams that run them.

K

Kill switch

A tested way to halt an AI system quickly.

Knowledge source

A body of content (documents, pages, databases) that an AI assistant or agent retrieves information from.

L

Layer

A grouping of domains used to sequence work: foundations, ways of working, practice.

Legal hold

An instruction that prevents deletion of content relevant to a legal matter or inquiry.

Lifecycle

The stages a use case passes through from intake to retirement.

M

Maturity level

One of five stages of capability: Initial, Repeatable, Defined, Managed, Optimising.

MCP (Model Context Protocol)

An open protocol that connects AI assistants and agents to external tools and data sources.

Moderation

The session in which assessors agree each score against its pass tests and evidence.

N

Near miss

AI behaviour that could have caused harm but was caught first, by a gate, a person or luck.

NIST (National Institute of Standards and Technology)

The United States standards body that publishes the AI Risk Management Framework.

O

OAIC (Office of the Australian Information Commissioner)

The national privacy and information regulator.

P

Pass test

The observable evidence test that decides whether a criterion has reached a level.

PDF (Portable Document Format)

A fixed-layout document file format.

PGPA Act

The Public Governance, Performance and Accountability Act 2013, which sets Commonwealth entities’ governance and performance duties.

Policy as code

Rules expressed in software so systems enforce them automatically rather than relying on people to remember them.

Priority score

Gap multiplied by importance and feasibility; used to order initiatives.

Prompt injection

An attack in which instructions hidden in content an AI reads (a document, email or web page) try to make it act against its instructions.

PROTECTED

An Australian Government security classification above OFFICIAL: Sensitive.

Provenance

The chain of custody of a decision or output: who or what produced it, from what, on what basis.

PSPF (Protective Security Policy Framework)

The Australian Government’s protective security policy, administered by the Department of Home Affairs.

Q

Quick win

An initiative that can be delivered in the first 90 days with modest effort and immediate value.

R

RACI

A table showing who is Responsible, Accountable, Consulted and Informed for each decision. Only one role may be accountable.

RAI (responsible AI)

Developing and using AI in ways that are safe, lawful, fair, transparent and accountable.

Realignment trigger

An event or result that requires the strategy or roadmap to be reviewed.

Red-team testing

Adversarial testing in which testers try to make a system misbehave, leak data or exceed its permissions.

Reference architecture

The approved technical design that AI workloads are built on, including the cloud landing zone, identity, network, logging and model access.

Register

The single list of AI use cases with owner, purpose, tier, lifecycle stage and assessment status.

Responsible AI overlay

A reading of the 40 criteria through five responsible AI dimensions, crosswalked to Australia’s AI Ethics Principles, ISO/IEC 42001 and the EU AI Act.

Retention class

A category of records with a defined retention period and disposal action, set by a records authority or policy.

Risk appetite

A written statement of how much AI risk the organisation will accept, by kind of risk, and what is outside appetite.

Risk tier

A category (for example low, medium, high) set by written criteria that decides how much assurance a use case receives.

Rollback

A tested way to return a system, and where possible its effects, to a previous state.

S

SAML (Security Assertion Markup Language)

A standard for single sign-on between an identity provider and an application.

SCIM (System for Cross-domain Identity Management)

A standard for creating and removing user accounts automatically from an identity provider.

Scope

Whether an assessment covers the whole enterprise or one business function.

Sensitivity ceiling

The most sensitive information (classification level, personal information) an AI tool is approved to handle.

Shadow AI

AI tools used for work without approval.

Shadow run

Running a new or changed AI system alongside current practice without letting it act, to compare its outputs before trusting it.

Simulation

Rehearsing an AI system in a safe environment under boundary, contradictory and adversarial conditions before deployment.

T

Target

The level a domain or criterion is meant to reach, with an ambition and a time frame.

Three lines of assurance

First line: owners operate controls. Second line: risk, privacy and security functions oversee and test them. Third line: internal audit assures them independently.

Transparency statement

A published statement describing how the organisation uses AI. Required for Australian Government agencies under Policy v2.0.

U

UHS Logic

An Australian technology consultancy and co-author of this model, which uses it in assessor-led engagements.

W

Workstream

A group of related initiatives with a lead role and a playbook.