AI in Finance in Aotearoa New Zealand: A Living Whitepaper

AI in New Zealand finance is moving into bounded production: customer service, fraud, claims and open banking are advancing, while advice, lending and agentic payments remain tightly governed, unevenly measured and largely human-supervised.

Executive Summary

• The strongest evidence of adoption remains in customer service, fraud prevention, workflow automation and claims support. Kiwibank has newly disclosed AI-powered customer-service tools, while Vero and Tower have reported live deployments in insurance operations. (kiwibank.co.nz)
• Open banking has moved from implementation to measurable use. MBIE reports more than 408,000 regulated open-banking payment requests worth NZ$130 million in July 2026, alongside more than 19 million data-sharing requests. (mbie.govt.nz)
• The open-banking operating model is changing. Payments NZ has announced that its API Centre will close at the end of September 2026, with standards-management responsibility transferring to MBIE. (paymentsnz.co.nz)
• Financial advice is now the FMA’s most explicit AI-specific supervisory focus. Its August thematic review is designed to understand how AI is being used, what conduct risks arise, and what safeguards firms have implemented. (fma.govt.nz)
• Evidence of autonomous consequential decision-making remains limited but is no longer absent. Tower reports completing one fully automated motor claim from lodgement through approval and payment without manual intervention. This is significant, but it is not evidence of scaled autonomous claims handling. (tower.co.nz)
• Agentic finance remains primarily preparatory. Visa’s Agentic Ready programme and Mastercard’s authenticated transaction with Westpac are testing the identity, consent, tokenisation, liability and fraud-control layers needed for agent-initiated payments. (visa.co.nz)
• The data is still fragmented and often vendor-reported. Experian research suggests substantial interest in AI-assisted underwriting among New Zealand institutions, but the sample is small and does not establish sector-wide production adoption. (ecommercenews.co.nz)

The sector’s direction is therefore clearer than its aggregate maturity: financial institutions are deploying AI where outputs can be bounded, reviewed and embedded into existing controls. The shift toward independent decision-making is visible mainly in isolated claims automation, advice experimentation, credit-risk tooling and payment demonstrations.

What Has Changed Since the Last Update

Open banking now has usage data, not just readiness claims

The previous article correctly identified open banking as maturing infrastructure. The latest MBIE figures strengthen that conclusion by showing actual regulated usage.

In July 2026:

• More than 408,000 open-banking payment requests were processed.
• The combined value of those payments exceeded NZ$130 million.
• More than 19 million requests to securely share banking data were processed.
• The number of accredited requestors reached 14, six more than in May.
• Three new accredited requestors—Centrapay, Experian and Trail—were added during July. (mbie.govt.nz)

These figures demonstrate live activity, although they should not be treated as evidence that AI-enabled financial products have already achieved mass adoption. Open banking is an enabling layer. The extent to which the data is being used for AI-driven affordability assessment, financial guidance, fraud detection or personalisation is not publicly quantified.

Stewardship of open-banking standards is moving to MBIE

On 21 August 2026, Payments NZ announced that the API Centre would cease operating at the end of September. Responsibility for standards management will transfer to MBIE. Payments NZ reported that more than 401,000 open-banking payments were completed in July, with more than 221,000 customers authorising payments or data-sharing consents during the month. (paymentsnz.co.nz)

This is an institutional change rather than an AI deployment. Its importance lies in the transition from an industry-coordinated implementation phase to a more formal regulatory and administrative model. For financial institutions and fintechs, future priorities will include continuity of standards, onboarding of accredited requestors, operational reliability and the expansion of services beyond the initial banking channels.

Kiwibank has disclosed AI-powered customer-service tools in production

Kiwibank’s financial results for the year ended 30 June 2026, released on 20 August, state that the bank introduced AI-powered tools within customer-service operations so staff could spend more time helping customers. The announcement does not disclose the specific tools, deployment scale, model type or performance outcomes. (kiwibank.co.nz)

This is nevertheless useful new evidence. It confirms that AI adoption is not confined to the largest internationally connected banks and insurers. It also reinforces the prevailing New Zealand model: AI is being introduced inside service operations to support employees, rather than presented as a replacement for frontline banking staff.

Kiwibank had already been using AI-enabled voicebot and customer-service capabilities through Genesys Cloud. Genesys reports reductions in transfers, average handling time and abandonment rates, but those results are vendor-reported and relate to an implementation that predates the latest financial-results disclosure. (genesys.com)

The previous MBIE consultation deadline has changed

The previous article stated that MBIE’s capital-markets consultation closed on 25 August 2026. The current MBIE page lists the submission deadline as 15 September 2026 at 5pm. The consultation continues to ask how New Zealand should respond to digital and AI-related innovation in financial products and markets. (mbie.govt.nz)

The consultation is not an AI-specific regulatory framework and does not authorise new AI use cases. Its significance is agenda-setting: AI is now explicitly part of the Government’s consideration of capital-market competitiveness, product design and market integrity.

Westpac has reached the announced deployment window, but outcome evidence remains thin

Westpac announced in April that it expected to deploy Microsoft Dynamics 365 Contact Centre as a Service across all its contact centres by August 2026. The platform gives customer-service staff real-time access to customer and product information during conversations. (westpac.co.nz)

The scheduled deployment window has now passed. However, the public information reviewed for this edition does not include official post-deployment metrics such as resolution time, customer satisfaction, error rates, escalation rates or workforce effects. The deployment should therefore be classified as an announced and apparently operational technology rollout, but not yet as a publicly evidenced productivity success.

The “no autonomous claims decisions” assessment needs qualification

The previous article assessed public evidence of autonomous consequential decision-making as limited. That remains broadly correct, but it requires a qualification.

Tower’s 2026 half-year results state that the company completed its first fully automated motor claim, from lodgement to approval and payment, without manual intervention. Tower also says it expanded automation across the claims process. (tower.co.nz)

This is the clearest publicly disclosed New Zealand example of a consequential insurance process being completed without manual intervention. It is not evidence that Tower has shifted to autonomous claims adjudication at scale: the disclosure describes a first completed claim, not the volume, eligibility rules, exception rates or governance arrangements for the wider system.

Current State of AI Adoption

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

Use case	Current position	Assessment
Fraud and scam prevention	A major regulatory and institutional priority, supported by real-time monitoring and detection systems	Established, but unevenly disclosed
Customer-service assistance	Live deployments at banks and insurers, including Kiwibank, Westpac, Tower and Vero	Most visibly scaling
Voicebots and self-service	Operating in selected banking and insurance contact centres	Mature in bounded workflows
Complaints and vulnerability detection	Vero is using sentiment analysis to identify dissatisfaction, complaints and possible vulnerability	Early production
Claims automation	Tower reports one fully automated motor claim and broader process automation	Early, tightly bounded
Financial advice	Digital tools, adviser copilots and some pure-AI advice examples are present; the FMA is investigating the sector	Exploratory and supervised
Credit and underwriting	Strong interest in AI-assisted decision support, but limited public evidence of autonomous lending at scale	Emerging
Open banking	Live payment and data-sharing activity is growing rapidly	Infrastructure operational
Agentic payments	Demonstrations and ecosystem programmes are underway	Experimental
Autonomous investment or trading	Little strong public evidence of scaled local deployment	Limited disclosure

Customer service is the clearest production pattern

The best-documented deployments place AI between a customer interaction and a human employee.

At Vero, Salesforce AI Sentiment Analysis was piloted across approximately 45,000 interactions before being rolled out across consumer and business operations. Vero says the system analyses calls and emails for sentiment, possible complaints and indicators of vulnerability, with staff expected to review flagged interactions and decide what action is appropriate. (suminsured.vero.co.nz)

Tower’s Amazon Connect deployment uses real-time transcription, automated quality assurance, call summaries, knowledge assistance and other AI-supported capabilities across sales, service and claims. Tower reports that customer interaction time fell by approximately 15% over seven months, with more than 796,000 minutes saved and average handling time reduced by two minutes and 38 seconds. (tower.co.nz)

A separate AWS case study reports different figures for a six-month period, including a 26% improvement in average call-handling time, an 18% improvement in email-handling time, 8,200 agent hours freed and a four-point increase in net promoter score. These figures are vendor- or company-supplied and use different measurement periods, so they should not be directly compared with Tower’s own announcement. (aws.amazon.com)

The common pattern is more important than the exact percentages: AI is being integrated into live service workflows where its outputs can be reviewed, corrected and audited.

Fraud and scam prevention remains strategically important

Fraud detection is one of the most consistently cited finance-AI use cases in New Zealand. The FMA’s 2026/27 Financial Conduct Report identifies fraud detection and prevention—including mortgage fraud, insurance fraud and fraudulent KiwiSaver first-home withdrawals—as a cross-sector priority. (fma.govt.nz)

Banks are also strengthening related controls through real-time transaction monitoring, identity verification, Confirmation of Payee and scam-intervention measures. However, public disclosure generally describes capabilities rather than performance. There is limited independently verified information on false-positive rates, fraud losses avoided, customer friction or the treatment of vulnerable customers.

This matters because fraud models often operate in consequential settings. A system that blocks or delays a payment may reduce fraud while also creating financial harm, distress or access problems for legitimate customers. The operational question is therefore not simply whether a model detects more suspicious activity, but whether institutions can explain, review and remediate its decisions.

AI is beginning to enter claims decisions

The Tower example indicates that AI and automation are moving closer to the decision boundary in insurance.

Tower’s AI-enabled contact centre is principally an assistive system. Its fully automated motor claim is different because it covers the complete journey from lodgement to payment. The available disclosure does not state whether AI made the underlying coverage decision, or whether deterministic rules and integrated assessing systems performed most of the work. It also does not disclose the claim’s complexity, value or eligibility criteria. (tower.co.nz)

The appropriate assessment is therefore bounded straight-through processing, rather than general autonomous claims adjudication. The case is strategically important because it shows that the boundary between workflow automation and consequential decision-making is beginning to move.

Advice remains predominantly human-enabled

The FMA’s March review of access to financial advice found examples of AI agents, compliance tools, client-analysis dashboards, record-keeping systems, budget-analysis tools and financial-wellbeing applications. It reported that the predominant use of AI was to enable advisers rather than replace them, with adviser oversight and accountability retained. (fma.govt.nz)

The same review found consumer trust in AI-provided advice varied from 28% to 41% across the products tested. Those results should not be interpreted as a national measure of consumer acceptance, but they indicate a substantial difference between willingness to use AI for administration or information retrieval and willingness to rely on AI for personalised financial decisions. (fma.govt.nz)

Lending and underwriting show high interest but limited local evidence

Experian research reported that 76% of surveyed New Zealand financial institutions were using agentic AI to support underwriters. It also reported that only 2% considered their data fully ready for AI-driven decision-making, while 69% described it as not ready or only partially ready. The New Zealand sample comprised 51 respondents. (ecommercenews.co.nz)

These findings are useful as an indication of market sentiment and experimentation, but they should not be treated as a national adoption statistic. The research is vendor-sponsored, the sample is relatively small, and “using agentic AI to support underwriters” may include pilots, decision support and limited workflow assistance rather than autonomous lending.

The broader conclusion is credible: appetite for AI in credit and fraud risk is ahead of the data, integration and governance foundations needed for dependable deployment.

Governance, Policy and Regulation

The FMA has moved from general research to use-case-specific supervision

The FMA’s 2024 research found that all 13 participating financial-services organisations either used generative AI or expected to adopt it soon. The research covered banking, insurance, asset management and financial advice, and found that firms were taking a cautious approach focused on security and risk management. (fma.govt.nz)

The regulator’s position is now more specific. Its 2026/27 Financial Conduct Report identifies digitisation, including AI, as an opportunity to improve access to advice, while also announcing a thematic review of how AI is used in practice, the associated conduct risks and the safeguards firms have in place. (fma.govt.nz)

The FMA published the thematic review on 6 August 2026. It is exploratory and fact-finding rather than a new set of licence conditions. Its immediate importance is evidential: it should give the regulator a better picture of whether AI is being used for administration, advice preparation, personalised recommendations, client interaction or final advice outcomes. (fma.govt.nz)

Prudential concerns extend beyond model accuracy

RBNZ’s May 2026 Financial Stability Report identifies several AI-related risks:

• Dependence on a small number of third-party AI providers.
• Biased, misleading or fraudulent outputs.
• Increased cyber risk.
• Operational-resilience weaknesses.
• Possible credit and employment effects from AI-driven disruption.
• Amplification of loss of confidence through social media and AI. (rbnz.govt.nz)

The Reserve Bank’s 2026 smaller-bank stress-test scenario includes a “name crisis” in which social media and AI contribute to deposit withdrawals. This does not mean that AI has caused such a run in New Zealand. It shows that supervisors now regard AI-generated information, synthetic content and automated amplification as part of the financial stability threat environment. (rbnz.govt.nz)

Agentic systems require a different control model

The National Cyber Security Centre’s May 2026 guidance on agentic AI recommends tight permissions, careful risk assessment and avoiding broad or unrestricted access to sensitive data or critical systems. It emphasises that agents can introduce risks through integrations, downstream actions, prompt injection and limited visibility into what the system has done. (ncsc.govt.nz)

For financial institutions, this has direct implications for payment agents, customer-service agents and internal workflow agents. The control question shifts from “Was the model’s answer accurate?” to a wider set of questions:

• What systems can the agent access?
• What actions can it take?
• Who authorised those actions?
• Can the action be reversed?
• Is there a complete audit trail?
• What happens when the agent encounters ambiguous or malicious information?

Privacy disclosures are becoming more explicit

ANZ updated its privacy statement in August 2026 to explain that AI may be used to support operations, customer service, complaints, fraud and scam detection, identity verification, credit and lending activities, and personalisation. It also stated that information may be analysed to create new insights. (anz.co.nz)

This is not evidence that every listed use case is operating in production. It is evidence that AI is becoming part of the formal customer-information and privacy architecture of a major bank. Such disclosures are likely to become more common as firms move from experimentation to embedded use.

New Zealand’s regulatory posture remains principles-based

The current framework is not a single AI law for finance. Instead, existing conduct, privacy, outsourcing, operational-resilience, prudential and fair-treatment obligations are being applied to AI-enabled systems.

MBIE’s responsible-AI guidance advises businesses to retain human review for outputs affecting customers, pricing, eligibility or finances, and to use agents initially for low-risk tasks with tight permissions. (business.govt.nz)

The FMA’s advice review, the RBNZ’s financial-stability analysis and the NCSC’s agentic-AI guidance all point in the same direction: firms can innovate, but they remain accountable for the outcomes produced by systems they deploy or procure.

Case Studies

Vero: sentiment analysis for complaints and vulnerability

Deployment: Salesforce AI Sentiment Analysis across consumer and business sales and service operations.

Function: Analyses customer calls and emails for positive, neutral or negative sentiment, possible complaints and indicators of vulnerability.

Reported scale: Approximately 45,000 interactions in the pilot; more than 65,000 sentiment outcomes after launch; approximately 10% of customer email and voice interactions analysed in real time.

Human oversight: Staff review flagged interactions and decide whether intervention is required.

Assessment: Vero provides one of the strongest New Zealand examples of AI being linked to conduct and customer-protection objectives. The system expands quality monitoring beyond small manual samples, but the reported figures are company-supplied and do not establish accuracy, bias or customer-outcome improvements independently. (suminsured.vero.co.nz)

Tower: AI-enabled service and a first fully automated motor claim

Deployment: Amazon Connect contact-centre platform with real-time transcription, AI-assisted support, knowledge assistance, summaries and automated quality assurance.

Reported outcomes: Tower says customer interaction time fell by approximately 15% over seven months, with more than 796,000 minutes saved and average handling time reduced by two minutes and 38 seconds. (tower.co.nz)

Claims automation: Tower’s half-year results state that the company completed its first fully automated motor claim from lodgement to approval and payment without manual intervention. (tower.co.nz)

Assessment: Tower shows two different adoption stages. Contact-centre AI is operating at meaningful scale, while end-to-end claims automation remains an early and bounded deployment. The claim example warrants close attention because it represents a move from employee assistance toward automated execution of a customer-affecting process.

Kiwibank: AI tools inside customer service

Deployment: Kiwibank’s FY26 results state that AI-powered tools were introduced within customer-service operations.

Related capability: Kiwibank has also deployed Genesys Cloud voicebot and contact-centre capabilities. Genesys reports reductions in transfers, abandonment, average speed of answer and average handling time, alongside improved agent-routing accuracy. (kiwibank.co.nz)

Assessment: The latest Kiwibank disclosure is important because it confirms continuing AI adoption at a New Zealand-owned challenger bank. The absence of technical detail and independently verified metrics means the maturity of the newer AI tools cannot yet be assessed. The operating model nevertheless appears consistent with the sector-wide pattern of AI-enabled service rather than autonomous banking decisions.

Westpac: AI-supported human conversations

Deployment: Microsoft Dynamics 365 Contact Centre as a Service, introduced in April 2026.

Function: Provides customer-service staff with relevant customer and product information during live conversations.

Status: Westpac announced an intention to complete deployment across all contact centres by August 2026. (westpac.co.nz)

Assessment: Westpac is a significant case because AI is being placed directly into regulated customer interactions at a major bank. The available public evidence confirms the rollout plan but does not yet provide official post-deployment results. It should therefore be treated as operational deployment with limited outcome disclosure, rather than a demonstrated productivity success.

Agentic payments: Mastercard, Westpac and Visa’s banking network

Mastercard and Westpac: In February 2026, Mastercard and Westpac completed authenticated agentic transactions in New Zealand using a Westpac-issued debit card to purchase cinema tickets. (mastercard.com)

Visa programme: ANZ New Zealand, ASB, BNZ and Kiwibank joined Visa’s Agentic Ready programme, which focuses on tokens, identity, risk and consumer controls for agent-initiated payments. (visa.co.nz)

Assessment: These are demonstrations and ecosystem-preparation activities, not mainstream agentic banking. Their significance lies in testing the trust layer around machine-initiated financial activity: delegated authority, transaction visibility, authentication, fraud monitoring, customer control and liability.

Trends

1. Production adoption is concentrated in bounded workflows

The most credible deployments have clear inputs, constrained outputs and identifiable human owners. Examples include transcription, summarisation, knowledge retrieval, sentiment classification, fraud alerts, call routing and workflow prioritisation.

This is a more conservative adoption pattern than the language of “autonomous finance” suggests. In New Zealand, firms appear to be using AI where it can improve throughput without transferring final responsibility for lending, advice, claims or investment decisions to a model.

2. Customer protection is becoming a practical route to adoption

AI used to detect vulnerability, identify complaints, prevent scams or improve service quality can be justified through customer outcomes as well as efficiency.

Vero’s deployment is particularly notable because it turns customer-interaction data into a conduct-monitoring signal. The risk is that sentiment or vulnerability classifications could themselves be inaccurate or overly simplistic. Strong governance will require escalation pathways, human review and monitoring for unequal treatment.

3. Data foundations are now the limiting factor

Open banking is providing structured, consented data-sharing infrastructure, while banks continue to modernise core platforms and customer-data environments. Yet vendor research indicates that data quality, lineage, integration and readiness remain substantial barriers to AI-driven decisioning. (ecommercenews.co.nz)

This creates a two-speed market. Firms can deploy relatively narrow AI features on top of existing systems, but more ambitious applications—such as real-time affordability assessment, personalised financial guidance or automated underwriting—depend on clean data, consistent definitions, reliable APIs and auditable decision records.

4. Agentic finance is developing through payments before advice or lending

Payments provide a relatively controllable environment for testing agent autonomy. Transaction limits, tokenisation, authentication and dispute processes can be defined more clearly than the boundaries of open-ended financial advice or credit decisions.

The current New Zealand activity therefore indicates infrastructure preparation rather than a commercial shift to autonomous finance. Adoption will depend on whether customers understand what an agent is authorised to do and whether institutions can allocate responsibility when an agent makes a mistake.

5. Vendor platforms are becoming part of the operating model

Many deployments are built around large technology platforms: Salesforce, Microsoft Dynamics, Amazon Connect and Genesys Cloud. This can accelerate implementation, but it also increases dependence on third-party providers, model updates, data-processing arrangements and platform availability.

RBNZ’s concern about concentration among third-party AI providers is therefore relevant even when the immediate use case is low risk. A small service tool can become operationally important if it is embedded across contact centres, claims processes or fraud operations. (rbnz.govt.nz)

6. Measurement and disclosure remain uneven

Public announcements commonly provide:

• Deployment dates.
• Customer or employee counts.
• Time saved.
• Handling-time reductions.
• Productivity percentages.
• Customer-experience scores.

They less often provide:

• Error rates.
• False-positive rates.
• Model drift.
• Bias testing.
• Complaint volumes involving AI.
• Human override rates.
• Privacy incidents.
• Workforce effects.
• Cost of operation.

This makes cross-firm comparison difficult. Some reported results are also not directly comparable because they cover different periods, populations or definitions of handling time. Tower’s differing company and vendor figures illustrate why headline performance claims should be treated cautiously. (tower.co.nz)

7. Financial advice is likely to become the next major governance test

Advice combines personal data, suitability, explanation, professional accountability and consumer trust. It is therefore more difficult to automate safely than call summarisation or transaction classification.

The FMA’s thematic review should clarify whether firms are using AI for internal productivity, adviser support, digital guidance, personalised recommendations or the delivery of regulated advice itself. Its findings may become the most important near-term reference point for the sector.

Outlook

Through the remainder of 2026

Several developments will determine whether the sector’s current AI activity becomes durable adoption:

1. FMA findings on AI in financial advice
   The regulator’s fact-finding work should reveal how far AI has moved into advice production and what controls firms are using. The key distinction will be between tools that prepare or support advice and systems that independently generate regulated recommendations.

2. Further open-banking growth
   July’s activity figures show momentum, but the next test is the quality and usefulness of services built on the infrastructure. Usage will need to translate into products that customers understand and trust, not merely higher API traffic. (mbie.govt.nz)

3. A managed transition of standards stewardship
   The transfer from Payments NZ’s API Centre to MBIE should preserve implementation continuity while clarifying responsibility for standards, accreditation and future expansion. (paymentsnz.co.nz)

4. More evidence from claims automation
   Tower’s first fully automated motor claim may be followed by further straight-through processing. Stakeholders should seek data on eligibility, exception handling, customer appeals, human review and outcomes for vulnerable customers.

5. Controlled agentic-payment trials
   Visa, Mastercard, banks and merchants are likely to continue testing agent identity, consent, transaction controls and dispute resolution. These trials will be more informative if firms disclose failure modes and customer safeguards, not just successful demonstrations.

Into 2027

The most likely path is incremental expansion rather than a sudden transition to autonomous finance:

• AI copilots will become standard in customer-service and operations teams.
• Fraud and scam systems will become more integrated 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, subject to FMA expectations.
• Claims and lending automation will grow in narrow, rule-defined segments.
• Model-risk management, third-party oversight and operational resilience will become more formal board-level responsibilities.

The sector may also face greater pressure to demonstrate workforce effects. Productivity gains are not equivalent to headcount reduction, and public trust may be affected if AI adoption is perceived primarily as a mechanism for reducing access to human support.

Overall Assessment

As of 1 September 2026, AI in Aotearoa New Zealand finance is operationally established in selected service, fraud, workflow and insurance applications, while more consequential uses remain bounded, experimental or weakly disclosed.

The previous assessment remains substantially sound but needs three updates:

• Open banking has progressed from technical readiness to measurable regulated usage.
• Kiwibank and Tower provide further evidence that AI is operating in live customer and claims workflows.
• Tower’s first fully automated motor claim shows that autonomous execution has begun in a narrow setting, even though scaled autonomous decision-making remains unproven.

The sector’s dominant model is still:

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

The strategic issue is no longer whether financial institutions can deploy AI. They can. The harder questions are whether they can measure its effects, govern third-party dependencies, explain its outputs, protect customer agency and demonstrate that efficiency gains improve—not weaken—fair treatment, resilience and trust.