Executive summary

When AI helps an organisation make decisions about people, people need a way to question those decisions. A person should be able to reach someone who will listen, examine what happened and act if something is wrong. The organisation then needs to ask whether others have been affected. This paper explores those three responsibilities and offers leaders a simple test: could someone unfamiliar with your organisation get a fair second look without having to fight for it?

Imagine asking an organisation for help and receiving an answer that doesn’t seem right. Perhaps an important detail has been missed, or your circumstances don’t fit the form. You want to explain. But who will listen, and can they actually change the answer?

This is not a new problem. Most of us have encountered a process that seemed unable to make room for our circumstances. AI can make that experience harder to untangle. It might read a document, summarise an application or help decide which requests receive attention first. A staff member may approve the result without noticing what was missed. By the time someone questions it, the mistake may have travelled through several parts of the organisation.

That is why the ability to reconsider a decision matters as much as the ability to make it efficiently. For boards, executives and people running services, there are three questions worth following.

Will someone listen when I say something is wrong?

Consider a hypothetical request for urgent assistance. The applicant provides the right document, but an AI tool reads a date incorrectly. Their request goes into the routine queue rather than receiving urgent attention.

The applicant may have no idea that AI was involved. They simply know they are still waiting for help. They should not need to understand the technology, or diagnose its mistake, before someone takes their concern seriously.

A clear way to contact a person is a useful beginning. So is explaining what will happen next and when they can expect an answer. But a contact form alone is not enough if the people most likely to need help struggle with internet access, language or disability. The route needs to work for the people using the service, not just for the organisation receiving the requests.

Sometimes a member of staff will notice the problem first. They need permission to pause and ask questions rather than pass on an answer they do not trust. This is what making room for an exception means here: recognising that a case needs attention outside the ordinary process. It is not a licence to ignore the rules.

If the person affected asks for the decision to be reconsidered, a review or appeal may be the appropriate next step. The available rights and time limits depend on the service and the laws or contracts governing it. A helpful informal conversation should not replace a formal review right. The Commonwealth Ombudsman’s automated decision-making guidance provides a useful starting point for government agencies, including on explanations and review. [1]

The practical point is straightforward: people should not have to know the correct administrative term before someone helps them find the right place.

Can that person help put it right?

Being heard matters. But imagine explaining the misread date to a sympathetic staff member, only to discover they cannot change the queue, see the original document or reach anyone who can.

The conversation may have been kind. The person is still waiting.

A meaningful second look requires more than someone being available to listen. The reviewer needs the evidence, the ability to question the original decision and the authority to act, or a clear route to someone who has it.

In our example, that means looking at the document itself. Asking the AI to explain its earlier summary may simply repeat the mistake. The organisation needs to understand what information it received, how it was used and why the request ended up in the wrong place.

For decisions with serious consequences, the reviewer also needs enough independence to reach a different conclusion. What that requires will vary, including with the applicable rules. The essential question is whether the review is genuinely open to changing the answer, rather than defending it.

There may also be something useful to do while the facts are checked. Within its powers, an organisation might pause an adverse action or arrange temporary help. Urgency should influence the response. A person whose necessary service has been interrupted cannot always wait for a lengthy investigation.

And when a correction is promised, someone needs to check that it happened. Moving our applicant into the right queue is only helpful if the request reaches the people who can provide assistance. The Ombudsman’s complaint handling guidance makes this connection between a response, an appropriate remedy and follow-through. [2]

Will the organisation check whether it happened to anyone else?

Once the applicant has been helped, it would be easy to close the case. Yet one question remains: was theirs the only date the tool misread?

Perhaps it was an isolated error. Perhaps the same kind of document has been misunderstood repeatedly. Either way, the successful challenge has given the organisation a reason to look.

That search should not be limited to people who complained. Some may not know a mistake occurred. Others may lack the time, confidence or energy to pursue it. A quiet complaints inbox is not evidence that every decision was sound.

Someone therefore needs responsibility for checking other potentially affected cases and investigating the cause. The work should protect personal information, especially in a small regional community where details can identify someone even without a name.

There are two different things to put right. People already affected may need their decisions reconsidered or an appropriate remedy. The process may also need changing so that the mistake does not continue. Repairing one does not automatically repair the other.

NIST’s voluntary AI Risk Management Framework connects monitoring with feedback, appeal, recovery and changes to a system. [3] It is useful guidance, not an Australian legal entitlement to appeal or compensation. The organisation must establish its own duties and powers.

Paper 14 considered accountability when AI-supported decisions are repeated across an organisation. A challenge to one decision is an opportunity to test that accountability in practice. Paper 16 takes the next step: what a proportionate remedy looks like and how to know whether the organisation has learned.

A practical test

For a smaller organisation in Ballarat, none of this necessarily calls for a specialist appeals office. It does call for a workable answer to each of the three questions.

Before introducing AI into a service, walk through what would happen if it helped produce a wrong answer. Start as someone who knows nobody inside the organisation. Could you reach a person, have the evidence reconsidered and obtain a correction where one is warranted? Would anyone check whether others were affected?

BRAIN’s role is to help organisations ask those questions, not to decide individual appeals. The responsibility stays with the organisation providing the service.

The test is not whether a review policy exists. It is whether the person waiting for help can get another look, and whether that second look can make a difference.

About BRAIN

The Ballarat Region Artificial Intelligence Network (BRAIN) is a regional institution focused on understanding the impact of artificial intelligence across Ballarat and surrounding communities and strengthening the capability required to respond. Through research, publications and regional collaboration, BRAIN examines how AI can contribute to regional prosperity and wellbeing while helping institutions navigate its risks and wider consequences.

About the Author

Matt Bowd is Co-Founder and Chief Executive Officer of the Ballarat Region Artificial Intelligence Network (BRAIN). His work focuses on AI governance, institutional capability and the implications of artificial intelligence for regional organisations and communities.

Sources

[1] Commonwealth Ombudsman, Automated Decision-Making: Better Practice Guide, March 2025, pp. 44–47. Guidance for government agencies on explanations, review and remediation.

[2] Commonwealth Ombudsman, Better Practice Complaint Handling Guide, 2023, pp. 29–32 and 36. Remedies, communication, follow-up and monitoring improvements.

[3] NIST, AI Risk Management Framework Core, AI RMF 1.0 (2023), MANAGE 4.1. Voluntary guidance on monitoring, feedback, appeal and recovery.