Governance Insights Series - Paper #08: From AI Policy to AI Assurance
Executive Summary
An AI policy can say which tools are approved, who is responsible and when human review is required. It cannot show that staff understand those rules, that material uses are visible or that a reviewer has time and authority to challenge an output. Assurance asks whether the governance described on paper operates in the work itself.
The practical chain is straightforward: identify an important claim, specify the control meant to make it true, obtain evidence that the control operates, review that evidence and act on what it reveals. If an organisation claims every consequential AI use has an owner, it should be able to name that owner, find the decision record and show who can change or stop the use. A declaration alone is weak evidence.
Assurance should be proportionate. A bounded drafting aid may need basic settings, staff guidance and periodic sampling. A system influencing care, employment or access to a service needs stronger testing, monitoring and independent challenge. The distinction follows consequence and reliance, not whether a supplier calls the product “AI.”
Regional organisations can build this practice through existing governance. The following paper examines the response when a failure is detected.
The Gap Between Policy and Practice
Policies have a necessary role. They establish expectations and help staff know what is permitted. Yet an organisation can prohibit sensitive data in public AI tools while employees remain unsure which information counts as sensitive. It can require a person to approve an AI-assisted decision while that person routinely accepts the output without examining it. It can demand vendor notifications but have no owner assigned to assess them.
These are gaps between a designed control and an operating one. Assurance does not assume a policy is false; it asks what evidence supports confidence that the policy is applied. A training completion report shows attendance, not understanding. A signed contract shows an obligation, not necessarily supplier performance. A register may list known systems while missing new features introduced through software updates.
Australia’s National Framework for the Assurance of Artificial Intelligence in Government treats assurance as part of broader governance across the AI lifecycle. The Commonwealth’s Policy for the Responsible Use of AI in Government, Version 2.0, requires accountability at agency and use-case levels within its defined scope. Regional organisations outside that scope should not present the policy as binding on them. They can still learn from its insistence that responsibility and use be made visible.
Assurance as a Decision Practice
Assurance is sometimes mistaken for a formal audit performed after deployment. Audit can be valuable for consequential uses, but leaders also need evidence while deciding whether to start, expand or continue a use. Assurance is the disciplined practice of obtaining sufficient, relevant information for that decision and exposing the limits of what is known.
Begin with a claim that matters. “Our staff review every AI draft before it goes to a resident” is more testable than “we use AI responsibly.” Management can define what review requires, sample real work and ask staff when they reject or alter an output. If the sample shows superficial checks, leaders can change the workflow, training or scope. Evidence has value because it can affect a decision.
The Digital Transformation Agency piloted an AI impact assessment tool across 21 Commonwealth agencies. Participants reported that it helped identify risks existing governance had missed, including fairness, transparency and explainability. That finding concerns the pilot and does not prove a particular tool will work in every institution. It illustrates the benefit of examining a use against its context instead of assuming ordinary controls capture every AI-related risk.
Know What Needs Assurance
An organisation cannot assure an AI use it has not recognised. AI may enter through a project, a purchased service, an upgrade to existing software or individual experimentation. A useful inventory identifies applications that can materially affect people, sensitive information, essential services, significant decisions or organisational resilience. It also records purpose, owner, status and supplier.
ASIC’s 2024 review of 23 financial services and credit licensees examined 624 AI uses deployed or in development and warned that governance risked lagging adoption. That is evidence from a particular regulated sample, not an estimate for Ballarat. Its lesson for other boards is to ask how management knows what is in use before it claims that all material uses are governed.
Make Accountability Observable
For each material use, a leader should be able to trace a decision. Who proposed it? Who considered the affected people and data? Who approved its purpose and limits? Who monitors it? Who can suspend it when assumptions fail? A high-level statement that “the executive is accountable” does not resolve these operational questions.
The record need not be elaborate. A small organisation may use a concise approval note and an existing risk meeting. The point is to ensure that the authority to act is understood by the people doing the work. Assurance might examine whether the owner actually receives reports, whether concerns reach a decision-maker and whether a vendor update triggers a fresh assessment when appropriate.
Test Whether Controls Work
Controls should be examined in the conditions under which staff use them. If the organisation requires human review, evidence might include a sample of decisions, documented changes to AI outputs and conversations with reviewers about time, skill and authority. If it restricts sensitive inputs, it can inspect configuration, guidance and examples of how staff classify information. If a supplier promises notice of material change, management can show how one such notice was handled.
Evidence must match the claim and potential consequence. A low consequence drafting task may need modest sampling. A use influencing eligibility for a service could require testing difficult cases, analysis of complaints, review of affected groups and independent specialist input. A single overall accuracy measure may conceal the errors most important to the institution.
The Office of the Australian Information Commissioner warns against a “set and forget” approach to commercially available AI products and recommends review of performance, training and monitoring through their use. This is privacy guidance with its own scope; the broader assurance lesson is that an approval at one point in time cannot establish continuing fitness for purpose.
Detect and Escalate Weak Signals
Assurance must help leaders see problems while they can still act. Error reports, staff overrides, complaints, unexpected output patterns, privacy concerns and service delays can each provide a partial view. A low count may reflect good performance or a reporting route nobody understands. Reviewing a signal requires context as well as a number.
Management should set thresholds appropriate to the use. A repeated minor drafting error may call for correction and sampling. A single serious error in a safety-related process may require an immediate pause and investigation. Staff and affected people need a route to raise concerns, and someone must have authority to respond.
Leaders can test the route before an incident by working through a plausible scenario. Would the right person hear about the problem? Could the organisation preserve records, stop the use and continue its service? Assurance examines the readiness of those arrangements. Paper No. 9 considers the response once something has gone wrong.
Reassess When Circumstances Change
AI systems and their settings can change after approval. A vendor may replace a model, a team may extend a tool to a new task, or staff may rely on an output more heavily than the original design assumed. Material change should trigger a review of the decision and the evidence supporting it.
The organisation can define what counts as material: a new purpose or data source, reduced human oversight, significant performance drift, a serious incident or a change affecting people’s ability to challenge an outcome. Routine updates need not go to a board or committee. What matters is that the boundary between routine maintenance and a changed governance decision is clear.
Proportionality in a Regional Institution
Smaller organisations should not copy the full machinery of a large regulator. They can integrate assurance into existing risk, privacy, service quality, procurement and board processes. The discipline is to identify the most consequential claims, find evidence that bears on them and put the results before someone able to act.
A regional institution might begin with a few material uses. For each, it records the outcome sought, accountable owner, principal risk, control, evidence source, review date and reason to escalate. It then tests whether those items can be located and whether the evidence changes anyone’s view. Missing evidence is itself a finding, not a reason to declare the use safe by default.
More consequential uses warrant more independent challenge. An operating team’s report may be supplemented by privacy, risk, clinical, legal or external expertise. Independence is useful when it asks a different question or tests an assumption that the team has reason to overlook. It should not become a ceremonial signature on an unchanged decision.
A Ballarat Scenario
Consider a hypothetical Ballarat education provider using an AI tool to draft feedback on student work. Its policy requires a teacher to review every draft. The provider wants assurance that the control works before extending the tool to more classes.
It could sample drafts and final feedback, examine where teachers corrected factual errors or tone, and ask whether review time is adequate. Students could explain whether the feedback is useful and how they challenge a mistake. If teachers rarely change the output, that may indicate excellent drafts or insufficient review; further inquiry is needed. Management can then decide whether to expand, alter training or limit the use.
The scenario describes no actual Ballarat provider. Its lesson is that assurance tests the practice behind a policy claim and leaves leaders with a decision to make.
Conclusion
AI policy establishes intention. Assurance examines whether the institution’s important claims are true in operation and remain true as use changes. It links a material use to an owner, a control, evidence, review and action. The evidence need not be perfect, but its limitations should be visible to those relying on it.
For regional organisations, this can begin small and become stronger with experience. A board or executive that can explain what it knows, how it knows it and what would change its decision has a sounder basis for confidence than one holding a policy alone. When the evidence signals failure, the question moves from assurance to response.
About BRAIN
The Ballarat Region Artificial Intelligence Network (BRAIN) is a regional institution focused on strengthening the capability required for communities and organisations to prosper in the age of artificial intelligence. Its Governance Insights Series offers practical, regionally grounded thinking for boards, executives and institutional leaders navigating AI-enabled change.
About the Author
Matt Bowd is CEO of BRAIN. His work focuses on AI governance, institutional capability and the implications of artificial intelligence for regional organisations and communities.
Sources
Digital Transformation Agency, Policy for the Responsible Use of AI in Government, Version 2.0.
Digital Transformation Agency, Pilot Findings Strengthen DTA’s Responsible AI Toolkit.
Next in the Series
BRAIN Governance Insights Paper No. 9 — Governing AI When Things Go Wrong. The next paper examines how an institution recognises an incident, decides who acts, responds to affected people and learns from the failure.
Written by Matt Bowd, CEO of the Ballarat Region Artificial Intelligence Network (BRAIN).
Each study is a step toward a more intelligent and resilient region.
To participate in regional pilots or insights partnerships, in our region or yours, connect via matt@brain.net.au.