AI in Finance

Executive Summary

• Operational adoption is expanding across customer service. ASB reports that conversational IVR helped reduce average call-wait times by about 40% in FY26. Kiwibank has introduced AI-powered customer-service tools, while Westpac, Tower and Vero have disclosed live AI-enabled service deployments. asb.co.nz
• Open banking has moved beyond implementation into measurable activity. MBIE reports more than 556,000 payment requests worth NZ$158.5 million and more than 24 million data-sharing requests in August 2026. mbie.govt.nz
• The open-banking operating model has now changed. The Payments NZ API Centre ceased operating on 30 September 2026. MBIE is taking responsibility for regulatory standards management, although the handover will continue over the following months. apicentre.paymentsnz.co.nz
• Financial advice remains the principal AI-specific conduct test. The FMA’s AI-in-advice surveys closed on 18 September 2026, but no findings had been published by 1 October. The available evidence still indicates that AI is used mainly to support advisers rather than replace them. fma.govt.nz
• Consumer trust is conditional rather than absent. Financial Services Council research found that 42% of consumers were comfortable with financial-services organisations using AI, but only 35% were comfortable with AI making automated financial decisions. Seventy percent expressed concern about privacy and personal data. These are industry-sponsored findings and should not be treated as an official national statistic. blog.fsc.org.nz
• Evidence of autonomous consequential decision-making remains narrow. Tower has reported one fully automated motor claim from lodgement to payment, while Sharesies’ AI Search is explicitly positioned as investment discovery rather than advice and warns that results may be inaccurate or hallucinated. earningscalls.dev
• The strongest evidence of scaled AI adoption is still in bounded workflows: contact-centre assistance, transcription, summarisation, sentiment analysis, fraud detection, customer triage and workflow automation. Evidence of scaled autonomous lending, advice, claims adjudication, investment management or payments remains limited.

The sector has therefore progressed since the previous update, but not in the direction of general-purpose autonomous finance. New Zealand institutions are adding AI to live processes while retaining human accountability, defined permissions and existing regulatory controls.

What Has Changed Since the Last Update

Open banking now has a second month of strong usage data

MBIE’s August figures reinforce the conclusion that regulated open banking is becoming an operating utility rather than a readiness exercise.

In August 2026:

• More than 556,000 open-banking payment requests were processed.
• The combined value exceeded NZ$158 million.
• More than 24 million banking-data sharing requests were processed.
• The number of accredited requestors remained at 14.

Compared with July, payment requests increased by 36%, payment value by 22% and data-sharing requests by 28%. The figures demonstrate live usage, although they do not show how much activity is being used for AI-driven affordability assessment, personalisation, financial guidance or fraud detection. Open banking is an enabling data and payments layer, not itself evidence of AI adoption. mbie.govt.nz

Payments NZ’s API Centre reported slightly different August figures—545,681 payments worth NZ$152.9 million—because its reporting covers its own ecosystem of standards users rather than the full MBIE regulated-activity measure. The difference illustrates why sector comparisons need to distinguish between regulatory reporting and industry-platform reporting. apicentre.paymentsnz.co.nz

The API Centre has closed, but open banking has not

The API Centre ceased operating on 30 September 2026, ending its industry-led role in open-banking standards development, management and governance. MBIE now leads the regulatory standards function under the Customer and Product Data Act 2025.

This is an institutional change, not an AI deployment. Its significance for AI lies in the infrastructure underneath future financial products: data access, consent, authentication, interoperability, reliability and auditability. The transition is not complete. Payments NZ says the handover of standards, tools, certification arrangements and support functions will continue after the closure date, while open-banking services remain operational. apicentre.paymentsnz.co.nz

ASB provides new quantified evidence from a live AI deployment

ASB’s FY26 results add one of the clearest bank-specific performance claims since the previous update. The bank says its conversational IVR can answer simple questions and route customers to the appropriate specialist, helping reduce average call-wait times by approximately 40% compared with FY25. asb.co.nz

This is stronger evidence than a generic statement that AI is being explored, but it remains a company-reported operational metric. ASB does not disclose the system’s containment rate, error rate, transfer rate, customer-satisfaction effects or how performance varies across customer groups.

The result strengthens the assessment that customer service is the leading production use case. It does not indicate that ASB has transferred responsibility for complex advice, lending or complaints decisions to an AI system.

The FMA’s AI review has passed its evidence-gathering deadline

The FMA opened its thematic review of AI in financial advice on 6 August 2026. It sought information from advice providers, technology firms, legal advisers and industry stakeholders on governance, oversight, suitability, record-keeping and consumer outcomes. The surveys closed on 18 September 2026. fma.govt.nz

No findings or sector-wide adoption figures had been published by 1 October. The review therefore remains a regulatory fact-finding exercise, not a new AI rulebook.

This qualifies but does not overturn the previous assessment. The FMA’s March access-to-advice review found that digital and hybrid advice models were expanding, while the predominant use of AI was to enable advisers rather than replace them. Regulatory returns also show that the estimated number of clients receiving digital advice rose from approximately 86,500 in 2024 to more than 164,800 in 2025. That growth is evidence of digital advice, not necessarily AI-generated advice. fma.govt.nz

RBNZ has placed AI within a longer-term banking-system framework

The Reserve Bank’s Future of Banking Study, published on 17 September 2026, considers three possible 2035 scenarios: continued dominance by incumbent banks, stronger digital challengers, and a more interconnected banking ecosystem.

The study does not predict which scenario will occur. It identifies AI, cloud services, data sharing and interconnected providers as forces that could increase efficiency and competition while also creating more complex operational, concentration and resilience risks. This moves the discussion beyond individual pilots toward the structure of the financial system itself. rbnz.govt.nz

Consumer-facing investment AI is now a clearer part of the evidence base

Sharesies continues to offer AI Search, which scans public financial data and produces investment-related summaries and search results. The platform explicitly says that AI Search does not provide financial advice, that no human has reviewed each result, and that outputs may be inaccurate, outdated or misleading. intercom.help

This is a useful boundary case. It shows that AI can be placed directly in front of retail investors without being presented as regulated personalised advice. It also demonstrates why disclosure, product classification and customer expectations matter: an AI tool can influence investment behaviour even when its provider says it is not making a recommendation.

Current State of AI Adoption

Public evidence points to meaningful operational adoption but limited transparency about aggregate scale.

Customer service is the clearest production pattern

The common design places AI between a customer interaction and a human employee.

At ASB, conversational IVR handles simple questions and triages calls. At Westpac, Microsoft Dynamics 365 Contact Centre as a Service provides staff with customer and product information during conversations. Kiwibank has disclosed AI-powered tools within customer-service operations. Tower uses real-time transcription, summaries, knowledge assistance and automated quality assurance across sales, service and claims. asb.co.nz

The reported benefits are mainly speed, reduced handling time and improved employee access to information. These are appropriate early use cases because outputs can be reviewed, corrected or overridden without necessarily changing the underlying product, price or eligibility decision.

The public evidence remains largely self-reported. Institutions generally do not publish error rates, escalation rates, repeat contacts, customer outcomes or the effect on vulnerable customers. Efficiency claims should therefore be treated as indicators of adoption, not independently verified proof of net benefit.

Fraud and scam prevention remains strategically important

Fraud detection, transaction monitoring, identity verification and scam intervention remain among the most credible uses for machine learning in finance. The FMA’s 2026/27 conduct priorities identify fraud detection and prevention as a cross-sector issue, while banks continue to invest in real-time monitoring, customer authentication and scam controls. fma.govt.nz

The evidence gap is significant. Public disclosures usually describe the existence of capabilities rather than their performance. There is little independently verified information on:

• False-positive rates.
• Fraud losses avoided.
• Customer friction created by interventions.
• Payment delays or blocked legitimate transactions.
• Outcomes for customers with limited digital access.
• How human reviewers handle disputed classifications.

The relevant question is not simply whether AI detects more suspicious activity. It is whether institutions can explain, review and remediate decisions that affect access to money.

Advice is digitally active but AI-specific evidence remains thin

The FMA’s evidence indicates growing interest in digital and hybrid advice, including adviser copilots, client-analysis dashboards, compliance tools, budgeting tools and financial-wellbeing applications. The predominant model remains adviser-enabled rather than adviser-replaced. fma.govt.nz

This distinction is important. A digital advice facility may use automated rules, forms and workflow software without using generative AI. Conversely, an adviser may use a generative AI tool internally without exposing it to the customer or relying on it for the final recommendation.

The FMA’s review should clarify where New Zealand firms sit across that spectrum. Until its findings are released, there is insufficient evidence to quantify production use of AI-generated recommendations, suitability assessments or personalised investment advice.

Lending and underwriting show demand ahead of public evidence

Experian research published in August reported that 76% of surveyed New Zealand financial institutions were using agentic AI to support underwriters. However, the New Zealand sample contained only 51 respondents, and the research was vendor-sponsored.

The same research found that 65% of respondents described their organisations as emerging or early in AI use across fraud and credit-risk underwriting. Forty-nine percent cited a lack of trust in AI outputs, 41% poor data quality and 37% fragmented data systems as challenges. Only 2% said they were comfortable with fully autonomous decisioning at scale across most use cases. ecommercenews.co.nz

These findings should not be treated as a sector-wide production-adoption statistic. They are more useful as an indication of market appetite and constraints. The available evidence suggests that lenders are interested in AI-assisted decisioning but remain cautious about delegating high-consequence credit decisions.

Claims automation has crossed a narrow threshold

Tower’s disclosure remains the strongest public New Zealand example of AI or automation reaching a consequential insurance process. The company reports completing its first fully automated motor claim from lodgement through approval and payment without manual intervention. It also reports broader claims automation and substantial contact-centre time savings. earningscalls.dev

This should be classified as bounded straight-through processing, not general autonomous claims adjudication. The public disclosure does not establish:

• How many claims are eligible for the automated pathway.
• The claim’s value or complexity.
• Whether AI made the coverage decision or supported a rule-based workflow.
• Exception and escalation rates.
• Customer appeal or remediation arrangements.
• Whether outcomes differ for vulnerable customers.

The strategic importance is nevertheless clear: a live customer-affecting process has been completed without manual intervention. The next evidence threshold is scale and outcome quality.

Investment discovery is live, but not advice

Sharesies’ AI Search shows a different route into financial services: AI is used to interpret natural-language queries, summarise public financial information and help users discover investments.

The product’s own warnings are significant. Sharesies says no human has considered the accuracy or relevance of each result, and that users should independently verify information before taking investment action. This makes the tool a live consumer-facing AI application, but not evidence of autonomous investment management or regulated personalised advice. intercom.help

Open banking and agentic payments are enabling layers

Open banking is now operating at meaningful volume, but public data does not identify how extensively AI is being applied on top of it. Potential uses include cash-flow analysis, affordability assessment, financial coaching, fraud detection and embedded payments. Evidence of those specific AI applications remains limited.

Agentic payments remain further behind. Mastercard and Westpac completed authenticated agent-led purchases in February 2026, while ANZ New Zealand, ASB, BNZ and Kiwibank joined Visa’s Agentic Ready programme. These activities test identity, consent, tokenisation, transaction visibility and controls. They are important technical and commercial demonstrations, but not evidence of mainstream autonomous payments. westpac.co.nz

Governance, Policy and Regulation

The FMA is becoming the central conduct reference point

The FMA’s AI-in-advice review is the most direct supervisory examination of AI use in New Zealand finance. It is exploring governance, oversight, suitability, record-keeping, consumer outcomes and legal questions across advice providers and technology firms. Its fact-finding status means firms remain subject to existing obligations rather than a new AI-specific licensing regime. fma.govt.nz

The FMA also assumed responsibility for the Credit Contracts and Consumer Finance Act from the Commerce Commission on 1 July 2026. That makes the FMA the single conduct regulator for financial markets, including consumer credit. The change is not AI-specific, but it increases the importance of FMA expectations for AI-assisted lending, affordability assessments, disclosures and fair treatment. comcom.govt.nz

Prudential concerns extend beyond model accuracy

RBNZ has identified AI-related risks involving third-party concentration, biased or misleading outputs, cyber risk, operational resilience, credit effects and the amplification of loss of confidence through social media and synthetic content. Its Future of Banking Study adds growing use of AI, cloud services, data sharing and interconnected providers to the long-term risk landscape. rbnz.govt.nz

For boards and senior management, this means AI risk cannot be treated solely as a model-validation issue. It also involves:

• Dependence on external model and cloud providers.
• Data-processing and privacy arrangements.
• Continuity if a provider, API or model becomes unavailable.
• Concentration across multiple institutions using the same platform.
• Cybersecurity and prompt-injection risks.
• Customer communication when an AI-enabled process fails.

Agentic systems require controls over actions, not only outputs

The NCSC’s joint guidance on agentic AI recommends careful risk assessment, narrow permissions and protection against threats arising through integrations, downstream actions and reduced visibility into what an agent has done. ncsc.govt.nz

For finance, the control questions are broader than whether a generated answer is accurate:

• What systems can the agent access?
• What payments, changes or communications can it initiate?
• What consent or authority limits apply?
• Can an action be reversed?
• Is there a complete audit trail?
• Can a human intervene before execution?
• What happens when the agent encounters ambiguous or malicious information?

These questions become especially important for payment agents, customer-service agents with account access and internal agents connected to lending or claims systems.

Open-banking governance is entering a regulatory-led phase

MBIE now manages the Consumer Data Right and is responsible for the regulated open-banking standards framework. The Ministry’s current guidance roadmap includes operational performance, outages, authentication, customer experience, testing and issue escalation. mbie.govt.nz

This is relevant to AI because future AI-enabled financial products will depend on reliable, permissioned and traceable data access. If the underlying APIs are inconsistent, difficult to authenticate or poorly understood by customers, more advanced AI use cases will inherit those weaknesses.

No standalone AI-finance regime has emerged

The policy picture remains principles-based. Existing conduct, privacy, credit, outsourcing, operational-resilience, prudential and cyber obligations continue to apply when AI is introduced.

MBIE’s capital-markets reform consultation, which included digital and AI-related innovation in financial products and markets, closed on 15 September 2026. No resulting AI-specific policy framework had been announced by 1 October. mbie.govt.nz

Case Studies

Tower: service automation moving toward straight-through claims

Tower’s Amazon Connect deployment uses transcription, summaries, knowledge assistance and automated quality assurance across customer interactions. The company reports a 15% reduction in customer interaction time over seven months and more than 796,000 minutes saved. These figures are company-reported. tower.co.nz

Tower also reports its first fully automated motor claim from lodgement to approval and payment without manual intervention. The case demonstrates that automated execution has begun in a narrow insurance workflow, but does not establish scaled autonomous claims decision-making.

Assessment: Strong evidence of production service automation; early evidence of bounded consequential automation; insufficient public evidence on scale, exceptions or customer outcomes.

Vero: AI sentiment analysis for complaints and vulnerability

Vero piloted Salesforce AI Sentiment Analysis across approximately 45,000 customer interactions before rolling it out across consumer and business operations. The system analyses calls and emails for sentiment, possible complaints and indicators of vulnerability. Flagged interactions are reviewed by staff. suminsured.vero.co.nz

Assessment: This is a notable use of AI for conduct monitoring rather than only efficiency. It could help identify dissatisfaction that manual sampling misses, but the public disclosure does not provide independent evidence of classification accuracy, bias or improved complaint outcomes.

ASB: conversational IVR in customer service

ASB reports that conversational IVR can answer simple questions and route callers to appropriate specialists, contributing to an approximately 40% reduction in average call-wait times compared with FY25. asb.co.nz

Assessment: A credible production use case with a quantified operational result. The result remains self-reported and does not show whether shorter waits translated into better resolution, fewer repeat contacts or improved outcomes for customers with complex needs.

Westpac: AI-supported human conversations

Westpac began rolling out Microsoft Dynamics 365 Contact Centre as a Service in April 2026. The platform provides customer-service representatives with relevant customer and product information during live conversations. Westpac expected full contact-centre deployment by August, but no public post-deployment performance metrics had been identified by 1 October. westpac.co.nz

Assessment: The deployment is evidence of live AI-enabled infrastructure, not yet of a publicly demonstrated productivity outcome.

Westpac has also become the first New Zealand bank accredited to issue digital identity credentials under the Government’s Digital Identity Services Trust Framework. This is not itself an AI deployment, but it is a relevant enabling development for identity verification, fraud reduction and future digital financial services. westpac.co.nz

Sharesies: AI Search for investment discovery

Sharesies’ AI Search generates investment-related summaries from public financial information and groups potentially relevant investments. The product is not presented as financial advice, and Sharesies warns that results may be inaccurate or outdated and have not been reviewed by a human. intercom.help

Assessment: This is one of the clearest examples of AI being placed directly in front of retail investors in New Zealand. It also demonstrates the practical boundary between an informational search product and a regulated personalised recommendation.

Mastercard and Westpac: authenticated agentic payments

Mastercard and Westpac completed authenticated agent-led transactions using a Westpac-issued debit card, including cinema-ticket and accommodation purchases. Visa has separately enrolled ANZ New Zealand, ASB, BNZ and Kiwibank in its Agentic Ready programme. westpac.co.nz

Assessment: These are controlled demonstrations and ecosystem-readiness activities. Their value lies in testing identity, consent, tokenisation, transaction visibility and liability before agents are permitted to act more broadly. They should not be described as mainstream autonomous banking.

Trends

1. Production adoption is concentrated in bounded workflows

The most credible deployments have defined inputs, constrained outputs and identifiable human owners. Transcription, call routing, summarisation, knowledge retrieval, sentiment classification and fraud alerts are easier to monitor than open-ended advice or lending.

New Zealand’s adoption pattern is therefore more conservative than the language of “autonomous finance” suggests.

2. The boundary is moving from assistance toward execution

The Tower claim and agentic-payment demonstrations show that AI and automation are moving closer to consequential actions. The change is incremental rather than wholesale:

• First, AI summarises or recommends.
• Then it prioritises or routes.
• Next, it executes within a defined workflow.
• Human review remains for exceptions, disputes and higher-risk cases.

The governance burden increases at each step.

3. Trust separates “helping” from “deciding”

The FSC research provides a useful demand-side distinction. Consumers are relatively receptive to AI that helps answer questions, identify scams or reduce paperwork. Comfort falls when AI determines whether a claim is paid, a recommendation is suitable, a loan is approved or an investment is appropriate. blog.fsc.org.nz

This suggests that transparency and human accountability will be commercial requirements as well as regulatory expectations.

4. Data foundations are limiting ambitious use cases

Experian’s survey points to poor data quality, fragmented systems and lack of trust in outputs as barriers to AI adoption. Open banking may improve access to structured, consented data, but it does not automatically solve data lineage, identity matching, affordability definitions or explainability. ecommercenews.co.nz

The likely result is a two-speed market: narrow AI features will continue to spread, while more ambitious decisioning applications remain constrained by data and governance.

5. Third-party platforms are becoming part of the operating model

Many public deployments use large technology platforms, including Microsoft, Salesforce, Amazon Web Services and Genesys. This accelerates adoption but creates dependencies around provider concentration, model changes, data residency, service availability and contract terms.

RBNZ’s concerns about cloud and third-party concentration are therefore relevant even when the initial use case appears low risk. A contact-centre tool can become operationally significant once embedded across customer service, complaints and claims. rbnz.govt.nz

6. Measurement remains less mature than deployment

Institutions increasingly disclose deployment dates, customer counts and time savings. They disclose much less about:

• Accuracy and hallucination rates.
• False positives and false negatives.
• Human override rates.
• Complaints involving AI.
• Bias testing.
• Model drift.
• Privacy incidents.
• Workforce effects.
• Total cost of ownership.

This limits meaningful cross-firm comparison. Vendor and company-reported metrics are useful signals, but they should not be treated as independently verified sector performance.

Outlook

Through the remainder of 2026

The following developments are likely to matter most:

1. FMA findings on AI in financial advice
   The key distinction will be between internal productivity tools, adviser support, digital guidance, personalised recommendations and AI that independently generates regulated advice.

2. The first full regulatory transition for open banking
   MBIE will need to maintain standards continuity, participant support, conformance processes and third-party engagement after the API Centre’s closure. Kiwibank is due to have account-information open-banking services ready from December 2026, which should broaden the data-holder ecosystem. mbie.govt.nz

3. More evidence from claims automation
   Stakeholders should look for disclosure of automated-claim volumes, eligibility rules, exception rates, appeals, human intervention and outcomes for vulnerable customers.

4. More bank-specific service metrics
   ASB’s 40% wait-time reduction is a useful benchmark, but the sector needs more information on resolution quality, repeat contacts, customer satisfaction and workforce impacts.

5. Controlled agentic-payment trials
   Visa, Mastercard, banks and merchants are likely to continue testing agent identity, consent, transaction limits, dispute resolution and fraud controls. The quality of evidence will improve if unsuccessful or disputed transactions are disclosed alongside successful demonstrations.

Into 2027

The most likely path is incremental expansion:

• AI copilots will become standard in customer-service and operations teams.
• Fraud and scam systems will integrate more closely with identity, payments and customer communications.
• Open-banking data will support more budgeting, cash-flow and embedded-finance services.
• Digital and hybrid advice models will expand under closer FMA scrutiny.
• Claims and lending automation will grow in narrow, rule-defined segments.
• Model-risk management and third-party oversight will become more formal board responsibilities.
• Customers will increasingly expect clear disclosure when AI materially shapes an interaction or decision.

The main strategic risk is not that New Zealand institutions fail to adopt AI. It is that adoption spreads faster than measurement, accountability and resilience controls.

Overall Assessment

As of 1 October 2026, AI in Aotearoa New Zealand finance is operationally established in customer service, fraud, workflow, sentiment monitoring and selected insurance processes, while advice, lending, investment decisions and agentic payments remain bounded, experimental or weakly disclosed.

Since the previous update:

• Open banking has continued its rapid growth and entered a regulatory-led transition.
• ASB has added quantified evidence of live AI-assisted customer service.
• The FMA’s advice review has completed its consultation phase but has not yet produced findings.
• RBNZ has placed AI within a broader analysis of how New Zealand banking could evolve through 2035.
• Consumer-facing investment discovery through AI is now a clearer part of the local evidence base.
• No evidence has emerged of scaled autonomous lending, advice, investment management or claims adjudication.

The dominant operating model remains:

AI as copilot, detector, classifier, workflow engine and bounded automation layer—with human accountability retained for higher-consequence outcomes.

The strategic question is no longer whether financial institutions can deploy AI. They can. The harder questions are whether they can demonstrate reliable customer outcomes, control third-party dependencies, protect privacy and agency, explain consequential decisions, and ensure that efficiency gains strengthen rather than weaken fairness, resilience and trust.