AI in Financial Services

Key takeaways

• ✓Australian banks and insurers are deploying AI across fraud detection, credit decisioning, claims processing and customer service, but the use cases carrying the most value also carry the most regulatory weight.

• ✓APRA and ASIC both have expectations around model risk, explainability and accountability that apply directly to AI systems. Getting the use case right and getting the governance right are not separate conversations.

• ✓The biggest risks in financial services AI are not technical failures. They are opaque decision-making, data quality problems and staff who do not know enough to catch errors before they cause harm.

• ✓Guardrails work when they are built into workflows, not bolted on after deployment. That means human review steps, audit trails, and clear ownership for every automated decision that touches a customer.

• ✓Training is a core part of the control environment. Staff who understand how AI tools work, where they fail, and what their obligations are under Australian law are a material risk reduction.

What use cases are banks and insurers actually deploying?

AI in financial services is delivering measurable value in four areas: credit decisioning, claims processing, fraud detection, and customer service. Each sits at a different point on the risk spectrum, which affects how much human oversight the deployment needs.

Credit and lending

Lenders are using AI to augment (not replace) credit assessment. Models trained on repayment history, transaction patterns, and bureau data can surface risk signals that a standard scorecard misses, particularly for thin-file applicants who lack a long credit history. The human credit officer still makes the final call on anything above a threshold, but AI flags anomalies and pre-populates the assessment file.

Document processing is the quieter win here. AI can extract income figures, employer details, and liability data from uploaded payslips and bank statements in seconds, cutting application processing time significantly without touching the credit decision itself.

Claims management

For insurers, the highest-volume opportunity is first-notice-of-loss triage. When a customer lodges a claim, AI can classify severity, check policy coverage, identify likely fraud indicators, and route the claim to the right handler before a human reads a single line. Straightforward motor claims with clean documentation can move from lodgement to settlement approval in hours rather than days.

Large language models are being used to draft settlement correspondence and coverage explanations, with adjusters reviewing and approving before anything goes to the customer. The adjuster writes less; they review more.

Fraud detection

This is where AI has the longest track record in financial services. Transaction monitoring models have been running in production at major banks for years, flagging unusual patterns in real time. What has changed recently is the quality of the explainability layer: compliance teams now expect the model to surface a human-readable reason for every flag, not just a risk score.

The newer challenge is authorised push payment fraud and voice cloning scams, where the customer is genuine but has been manipulated. AI detection here is earlier-stage and less reliable, which makes staff training as important as the technology.

One risk that training often misses

Fraud detection models catch unusual patterns, but social engineering attacks succeed precisely because the behaviour looks normal. Building AI literacy across customer-facing teams is part of the fraud control framework, not a separate initiative.

Customer service and operations

Generative AI assistants are handling a growing share of inbound queries: account balance enquiries, product FAQs, hardship application guidance, and complaint triage. The key design decision is knowing exactly where the assistant hands off to a human. Banks that deploy an AI assistant without a clean escalation path end up with frustrated customers and complaints to AFCA (the Australian Financial Complaints Authority).

Back-office operations are equally productive territory. Regulatory reporting, policy document summarisation, meeting notes, and internal knowledge search are all in production at Australian financial institutions. These carry lower customer risk, so they are often the right place to start building confidence before moving to customer-facing applications.

Where does AI create the most risk in financial services?

The highest-risk moments in financial AI are not spectacular failures. They are quiet ones: a model that consistently underscores loan applications from certain postcodes, a chatbot that invents a product fee, a summary that omits a material disclosure. Each of these can cause real harm before anyone notices.

Model bias in credit and underwriting

Machine learning models trained on historical data inherit the biases in that data. In lending and insurance underwriting, this matters because the consequences fall on real customers. A model trained on approval patterns from a period when certain demographics were systematically underserved will replicate that pattern at scale, faster and with more consistency than any human team.

Australian financial services firms also carry obligations under the Equal Credit Opportunity provisions and broader anti-discrimination law. If a model is producing disparate outcomes by postcode, age bracket, or other proxies for protected attributes, the firm holds the liability even if the model is supplied by a third party.

Hallucination in customer-facing outputs

Large language models (LLMs) generate plausible text. They do not verify it. In financial services, that distinction is critical. A general-purpose AI assistant asked about margin loan terms, superannuation contribution caps, or insurance policy conditions can produce a confident, well-formatted answer that is factually wrong.

The risk compounds when customers act on that output. The remediation path, whether that is a compensation payment, a regulatory breach notice, or a complaints handling process, is expensive. The ChatGPT for business article covers this pattern in more detail, but the short version is: LLMs should not be the source of truth for product terms or regulated disclosures without retrieval controls anchoring the output to verified sources.

Data privacy and confidentiality

Financial institutions hold some of the most sensitive personal data that exists: income, debt, health-linked insurance data, transaction histories. When staff use general AI tools without clear guidance, that data moves. It moves into prompts, into chat histories, into model fine-tuning pipelines if the tool's terms permit it.

The Privacy Act reforms introduced stronger obligations around automated decision-making and the handling of personal information. The practical implication for a bank or insurer is that every AI tool in use needs a data handling assessment, not just the ones IT procured formally. Shadow AI use by well-meaning staff is one of the more common sources of exposure.

The risk you are probably underestimating

Shadow AI use, staff copying client data into consumer-grade tools without malicious intent, is a more common source of Privacy Act exposure than deliberate data exfiltration. Training is cheaper than remediation.

Regulatory exposure from automated decisions

ASIC and APRA both expect that firms can explain how a decision was reached. If the answer is "the model decided," that is not an explanation. Explainability requirements apply to credit decisions, claims outcomes, and any output that materially affects a customer.

This is not just a compliance formality. When a customer disputes a declined claim or a credit refusal, the firm needs an audit trail. Black-box models, those where the internal reasoning cannot be surfaced, create legal and regulatory exposure precisely because they cannot be explained in a form that satisfies a regulator or a court. The four AI risks every board should be asking about covers the governance layer that sits above these operational risks.

Which regulations apply to AI in Australian financial services?

There is no single "AI Act" in Australia yet, but that does not mean financial services organisations are operating in a regulatory vacuum. Several existing frameworks already govern how AI can be deployed, and new guidance is arriving steadily.

APRA (Australian Prudential Regulation Authority) is the most directly relevant body for banks, insurers, and superannuation funds. Its prudential standard CPS 230, which addresses operational resilience, creates clear obligations around material service providers and technology risks. If your organisation is using an AI system to make or inform credit decisions, underwriting, or claims handling, that system is likely a material service provider arrangement under APRA's framework. APRA has also signalled that it expects boards to understand and govern technology risk, which extends to AI systems that affect financial outcomes.

ASIC (Australian Securities and Investments Commission) has published guidance on digital advice and automated decision-making. Its focus is on whether AI-assisted processes meet the same conduct obligations as human ones: best interests duties, disclosure requirements, and the obligation not to mislead consumers. An AI tool that generates financial advice, summarises product terms, or routes customer complaints needs to meet the same standards a human adviser would.

The Privacy Act governs how personal information is collected, used, and disclosed. AI systems in financial services frequently process sensitive personal data, including credit history, health information for insurance underwriting, and transaction behaviour. The Privacy Act's Australian Privacy Principles place obligations on how that data can be used as model training data, and whether inferences drawn from it constitute a new "use" of personal information. The 2024 amendments to the Privacy Act are particularly relevant here.

The AI accountability framework is still forming. The Australian Government released its interim response to the Safe and Responsible AI consultation in 2024, signalling a risk-tiered approach rather than broad legislation. Financial services sits in a category likely to attract mandatory guardrails given the potential for consumer harm. Organisations that build governance now are better placed when mandatory requirements land.

The gap between 'no AI law yet' and 'no obligations yet'

Existing prudential, conduct, and privacy obligations already apply to AI systems. Waiting for dedicated AI legislation before building governance is a risk posture, not a neutral one.

Practically, this means a bank or insurer deploying AI needs to map each use case against at least three questions. Does the system affect a regulated financial outcome? Does it process personal information? And does it create advice, recommendations, or disclosures that carry conduct obligations? The answers determine which frameworks apply and how urgently.

For a more detailed look at how these frameworks interact with day-to-day AI deployment, the sibling article on AI compliance for banks and insurers covers the governance layer in depth.

How do you build guardrails that actually work?

Guardrails fail when they are treated as a compliance checkbox rather than a design constraint. The teams that get this right build controls into their AI workflows from the start, not after something goes wrong.

Four areas matter most.

Human-in-the-loop design

For any decision with material consequences, a human must be able to review, override, or reject an AI output before it takes effect. This is not about distrust of the technology. It is about accountability: under Australian financial services law, your organisation remains responsible for the outcome regardless of how it was generated.

In practice, this means mapping your AI use cases to decision types. Routine, low-stakes outputs (summarising meeting notes, drafting internal communications) can run with minimal oversight. Credit decisions, claims outcomes, and customer-facing advice require a named reviewer and a clear audit trail.

Data classification

Not all data should flow into every AI tool. A prompt engineer at a retail bank discovered this the hard way when a staff member pasted a customer's full transaction history into a public-facing AI assistant for analysis. No malice, just a gap in training and tooling.

Before deploying any AI tool, classify your data by sensitivity: publicly available, internal, confidential, and regulated (which includes anything covered by the Privacy Act or the CDR). Then configure your tools to match. Enterprise deployments of ChatGPT, Claude, and Copilot all offer data handling controls, but they require deliberate configuration, they do not activate by default.

Output review and audit trails

AI outputs in financial services should be logged, not just acted on. This matters for two reasons. First, regulators may ask you to demonstrate how a decision was reached. Second, logging lets you spot model drift, where an AI tool that performed well at deployment starts producing outputs that diverge from your standards over time.

Build a lightweight review cadence into high-risk use cases: spot-check a sample of outputs weekly, flag anomalies, and feed findings back into your training and configuration. The four AI risks every board should be asking about includes model drift as a governance exposure that often goes unmonitored until it surfaces as a complaint or a breach.

Staff training as a control

Technology controls are only as strong as the people operating around them. A well-configured AI tool will still be misused if staff do not understand what the guardrails are, or why they exist.

AI compliance for banks and insurers is partly a people problem. Staff need to know which tools are approved, what data can be used in each, when to escalate an AI-generated output for review, and what to do if they suspect an error. That is not a one-hour induction topic. It requires scenario-based training specific to the workflows your teams actually run.

Guardrails are a system, not a document

A policy document that no one has read is not a guardrail. Effective controls combine technical configuration, clear decision rules, audit logging, and staff capability, applied consistently across every team using AI.

What does good AI training look like for a financial services team?

Generic AI training teaches people how to write prompts. That is useful, but it is not enough for a bank or insurer operating under APRA, ASIC, the Privacy Act, and the emerging obligations flowing from the EU AI Act's influence on Australian policy. Staff who know how to generate text but do not understand what can go wrong with customer data, credit decisions, or claims handling are a liability.

Sector-specific training covers three things generic programs typically skip.

The compliance context. Staff need to understand which AI outputs are considered "decisions" for regulatory purposes, what APRA's CPG 234 and CPG 235 say about technology risk, and when a human must review an AI recommendation before it is acted on. That framing changes how people use AI tools day to day.

Data handling in practice. A claims analyst who pastes a customer's medical history into a public-facing AI tool is not being reckless on purpose; they simply have not been shown what the boundary is or why it exists. Training needs to make data classification concrete, with examples that mirror the team's actual work rather than abstract privacy principles. The Privacy Act changes and what they mean for AI use are relevant reading here.

Workflow-level judgment. The highest-value skill in a regulated environment is knowing when to trust an AI output, when to verify it, and when to escalate. That is not something a one-hour onboarding video builds. It requires scenario-based training that puts people in situations where the AI is plausible but wrong, or where the output is accurate but using it would breach a policy.

The gap generic training leaves open

Most AI tools training teaches capability. Financial services teams also need training on constraint: what the tool cannot be used for, why, and what the consequence is of getting it wrong.

Role design matters too. A credit analyst, a fraud investigator, and a customer service team leader all use AI differently and face different risk profiles. A single all-staff session tends to be too broad to be actionable for any of them. Cohort-based delivery, where content is built around the decisions each role actually makes, produces faster adoption and fewer compliance incidents.

The same applies to seniority. Executives need a different conversation to frontline staff. Board members and senior leaders benefit from understanding AI risk at a governance level, including how to ask the right questions of their technology and data teams, while analysts and operators need hands-on proficiency with the tools their organisation has approved.

For organisations building out AI capability at scale, Better People's enterprise AI training programs are designed to run across multiple teams and cohorts, with content adapted to sector-specific regulatory requirements rather than applied generically.

Frequently asked questions

Is AI use in Australian financial services regulated right now?

No single AI-specific law governs financial services in Australia today, but several existing frameworks apply directly. APRA's prudential standards (particularly CPS 230 on operational risk and CPS 234 on information security) require regulated entities to manage technology risk, which includes AI systems. ASIC's responsible lending and market integrity obligations also apply wherever AI influences credit decisions or customer advice. The Privacy Act 1988 governs how customer data is collected, used and disclosed, including by AI tools. Expect more prescriptive AI guidance from both regulators in the near term; the Australian Government's AI Safety Standard consultation signals that direction clearly.

Which AI tools are safe to use with customer data?

No consumer AI tool is safe to use with identifiable customer data without specific configuration and contractual controls. Tools such as ChatGPT Enterprise, Microsoft Copilot with appropriate data loss prevention settings, or Claude for Enterprise can be configured to avoid training on your data and to enforce access controls, but the defaults on consumer versions do not offer those protections. The ChatGPT Enterprise data privacy settings are meaningfully different from the standard product. Before deploying any tool, your legal, privacy and information security teams need to review the data processing agreement, confirm where data is stored, and assess whether that meets your obligations under the Privacy Act and your APRA operational risk framework.

What staff training do banks and insurers actually need?

Staff need three things that most generic AI training does not cover: an understanding of which tasks are within acceptable use policy, the ability to recognise when AI output requires human review, and clear escalation paths when something goes wrong. Front-line staff and relationship managers need practical scenario-based training built around your actual tools and workflows. Compliance and risk teams need deeper exposure to model limitations, bias risks, and audit requirements. Senior leaders need enough fluency to set policy and ask the right questions of vendors. Training that covers only how to write a prompt misses most of what matters in a regulated environment.

How do we prevent AI from producing biased or discriminatory credit decisions?

Bias in credit decisioning is a live regulatory risk, not a theoretical one. AI models trained on historical data can encode past lending patterns that disadvantage certain groups, potentially breaching anti-discrimination law under the Australian Human Rights Commission Act and relevant state legislation, as well as responsible lending obligations under the National Consumer Credit Protection Act. Mitigations include regular bias audits across demographic segments, explainability requirements so declined applicants can be given genuine reasons, human review thresholds for high-stakes decisions, and clear documentation of how the model was trained and validated. Embedding these controls at deployment, rather than retrofitting them later, is substantially cheaper and far less legally exposed.

How quickly can a financial services team realistically build AI capability?

A meaningful baseline of AI fluency across a mid-sized team is achievable in three to six months with structured training, clear use case prioritisation, and leadership that actively models the behaviour. The limiting factor is rarely the technology. It is the absence of internal policies, unclear ownership of AI governance, and training that does not connect to real workflows. Teams that start with a small number of well-scoped, low-risk use cases, train staff on those specific tasks, and build governance as they go consistently outperform those that run a broad rollout without the supporting structure. See our enterprise programs for how we approach this with regulated organisations.

Ready to build AI capability that fits your regulatory environment?

Financial services teams don't need generic AI awareness training. They need programs that address the specific tools they're deploying, the compliance obligations they're working under, and the judgment calls their people face every day.

Better People designs and delivers AI training for banks, insurers, and other financial institutions across Australia. That means working through your actual use cases, your data classification requirements, and the guardrails your governance team has already set, so the training sticks in a regulated environment rather than simply describing what AI can do in theory.

See how we work with financial services teams or talk to us about what your organisation needs.