On 17 February 2026, in Auckland, an AI agent called Matilda made two purchases on behalf of an unnamed Westpac NZ customer. It bought cinema tickets from Event Cinemas. It booked accommodation at QT Hotels in Queenstown. Both transactions settled through Westpac-issued debit cards, processed by IPSI, and validated by Mastercard’s Agent Pay framework.
This was the first time an AI agent had completed an authenticated financial transaction in New Zealand. The model that did it was Maincode’s Matilda — built in Melbourne, Australia.
That’s the story. And it raises a question New Zealand’s tech sector should be asking loudly.
🔍 THE BOTTOM LINE
New Zealand’s first agentic transactions ran on an Australian sovereign LLM, processed by a local payment orchestrator (IPSI), and authorised by Mastercard’s global Agent Pay rails. The bank was Westpac NZ. The AI was Maincode’s Matilda. The model that did the work is Australian. This is a successful pilot. It’s also a sovereignty case study.
THE TRANSACTIONS
The Mastercard press release, dated 17 February 2026, lays out the two pilot transactions in detail:
| Detail | Transaction 1 | Transaction 2 |
|---|---|---|
| Type | Cinema tickets | Accommodation |
| Merchant | Event Cinemas (EVT) | QT Hotels & Resorts (Queenstown) |
| Payment method | Westpac-issued debit card | Westpac-issued debit card |
| Processor | IPSI | IPSI |
| AI agent | Maincode’s Matilda | Maincode’s Matilda |
| Cardholder consent | Full | Full |
| Authorisation | Fully authenticated | Fully authenticated |
Both transactions were visible to all parties — issuer, acquirer, merchant — and explicitly recognised as agent-conducted. No chargebacks, no fraud, no manual override required.
THE STACK
Four parties made this work. Three of them are Australian or global. One is New Zealand.
1. Westpac NZ — the issuer of the debit card, the regulated bank, the consumer-facing brand. The only party with the customer relationship.
2. Maincode (Melbourne) — built Matilda, the LLM that handled intent interpretation, merchant selection, and transaction negotiation. Australian-sovereign, not NZ.
3. IPSI (Auckland) — the local payment orchestrator, processing layer, and technology integration partner. New Zealand-owned, 20+ years operating in APAC.
4. Mastercard (global) — provided the Agent Pay framework, identity layer, fraud and AML controls, and the global standards interoperability.
Read that list again. The New Zealand components of New Zealand’s first agentic payment are: the bank and the payment processor. The AI model is Australian. The framework is American. The merchant is global.
🇳🇿 THE SOVEREIGNTY QUESTION
Maincode is a serious company. Matilda is a credible sovereign LLM, and their work with Mastercard shows real technical depth. Lukas Wesemann, who leads AI research at Maincode, is on the record saying: “Building AI systems that can not only reason but reliably act and facilitate payments is a fundamental shift in how business will be done online.”
That’s all true. But the question remains: why wasn’t Matilda a New Zealand model?
New Zealand has:
- World-class ML researchers (University of Auckland, Victoria, Canterbury)
- A national AI strategy (MBIE’s Blueprint for AI in Aotearoa)
- Active sovereign-AI investment via Callaghan Innovation
- Existing local LLM work (some at smaller scale, mostly in research)
The decision to use Maincode’s Matilda rather than build or commission a New Zealand model is not, on its face, a bad call. Maincode is one of the few sovereign-AI labs in Australasia with proven deployment at financial-grade reliability. Building a comparable New Zealand model from scratch would have taken 12-18 months and significant capital — a fact acknowledged in MBIE’s Blueprint for AI in Aotearoa when it flagged sovereign-LLM capability as a multi-year build.
But the longer-term question is whether New Zealand banks, fintechs, and government services are going to keep deferring to Australian and American sovereign AI for the most consequential applications — or whether NZ will fund and develop its own.
Sarah Hearn, Westpac NZ’s Managing Director of Product, said the bank is “investing heavily in innovative technology and processes to make banking faster, safer and easier for New Zealanders.” That’s the right corporate answer. It is not, however, a sovereign-AI strategy.
📊 WHAT NEW ZEALAND CONSUMERS ACTUALLY WANT
The Mastercard release cites Adobe consumer research, and the numbers are striking:
- 69% of NZ consumers are convinced agentic AI will make their lives easier
- 75% are appealed by AI that can interpret intent, make decisions, and take autonomous action
- 83% of NZ shoppers already using AI assistants report positive experiences
- Most popular agentic categories: Entertainment (36%), Clothing (34%), Health & Beauty (32%)
The demand side is not the problem. New Zealanders are ready. The supply side — local models, local agents, local infrastructure — is the bottleneck.
🏪 THE MERCHANT VIEW
Andrew Turner, Group General Manager Technology and Digital Transformation at EVT (which operates Event Cinemas, QT Hotels, and Rydges in NZ), was on the Mastercard release:
“Agent-powered commerce could reshape how New Zealanders discover, book and experience entertainment and travel, and EVT is proud to be part of this milestone with Mastercard. We are focused on making it effortless for guests to book a movie at Event Cinemas or plan a stay at a QT or Rydges.”
This is a significant statement. EVT is signalling to every AI platform, every payment processor, and every competitor in entertainment/travel that they intend to be agent-ready infrastructure for the New Zealand market. That’s a strategic commitment that goes beyond a single pilot, and it aligns with the broader trend Adobe’s consumer research captured when 69% of NZ consumers said agentic AI would make their lives easier.
🔐 THE TRUST MECHANISM
Mastercard’s Agent Pay framework is the linchpin. It’s the standards layer that made this work, and it’s worth understanding what it actually does:
- Identity verification — confirms the agent is acting with cardholder consent
- Transaction visibility — issuer, acquirer, and merchant all see the agent
- Fraud and AML — embedded natively in the workflow
- Dispute resolution — clear chain of accountability
- Interoperability — works across issuers, acquirers, and merchants regardless of platform
This is exactly what agentic commerce needs. Without this layer, agent payments are a regulatory nightmare and a fraud target. With it, they’re a payment method like any other.
THE GLOBAL PICTURE
Mastercard’s release explicitly notes this is part of a regional rollout. The New Zealand milestone follows:
- January 2026 — Australia (also via Maincode Matilda)
- Q1 2026 — UAE pilot
- Q1 2026 — Latin America pilot
- Q2 2026 — New Zealand
The implication: Mastercard is rolling out the same sovereign-AI agentic payments stack across multiple markets, with Maincode Matilda as a key infrastructure partner in the region. This is not a one-off NZ experiment. It’s a global architecture where the AI layer is being deliberately localised but not necessarily sovereign to the country of use.
❓ FREQUENTLY ASKED QUESTIONS
Is Maincode Matilda a good model?
Yes, by available evidence. It handled two real financial transactions with full authentication, processed by a regulated payment orchestrator, validated by Mastercard’s framework. The pilot succeeded. The technical questions are about scale, cost, and whether the model can handle the long tail of NZ consumer intent (which differs from Australian intent in non-trivial ways).
Could a New Zealand model have done this?
In principle, yes. In practice, no — not in February 2026. New Zealand lacks a sovereign LLM with proven financial-grade reliability, and building one takes time. The interesting follow-up question is whether NZ will have one by February 2027.
Is this a one-off, or is it happening at scale?
Mastercard’s stated plan is to scale agentic commerce across the region. The two NZ transactions were a proof-of-concept. The next 12-18 months will see real production volumes if the architecture holds.
What about x402 and the crypto-rail story?
Different stack, similar goal. Mastercard Agent Pay uses card networks. x402 uses stablecoin settlement. Both are racing to be the default agentic commerce rail. The first one to hit a billion-agent threshold wins the network-effect race.
Does this mean Westpac is ahead of ANZ, ASB, BNZ?
In agentic payments, yes — they’re the first NZ bank to publicly complete an authenticated agentic transaction. The other three are surely not far behind, and watching Mastercard’s regional rollout will tell us who’s next.
What does this mean for NZ consumers?
In the short term, almost nothing. You’ll still book your own cinema tickets. In 18-36 months, expect AI assistants to handle routine purchases, subscription management, and travel booking on your behalf — with Westpac, IPSI, and Mastercard as the rails. You won’t see “an AI did this” on your statement. You’ll just see a normal debit.
Should NZ be worried about relying on an Australian model?
The pilot worked. The dependency is real. New Zealand needs to decide whether the next generation of agentic infrastructure will be built locally, regionally, or imported. That decision has implications for Callaghan Innovation funding, university research priorities, and how NZ banks approach procurement.
❓ FAQ
Why did Westpac choose an Australian LLM and not a New Zealand one? There isn’t one ready. The Maincode Matilda sovereign LLM is the only Australian-made model of sufficient capability for financial transactions, and at the time of the February 17 pilot, no NZ-built model had completed certification for production banking use. The sovereign-AI question this raises is real: NZ banks are now depending on Australian infrastructure for cutting-edge financial services, while NZ invests nothing comparable in domestic LLM capacity.
Were customers protected if the AI made a bad purchase? Yes — under Mastercard’s Agent Pay framework, the transactions went through standard Westpac-issued debit cards, so the existing consumer protection regime (chargeback rights, fraud liability) applied. The customer could dispute the purchase the same way as any other card transaction. The novelty is the AI-initiated decision, not a regulatory gap.
Could the AI have made purchases the customer didn’t want? In theory yes, in practice no. The agent was operating within pre-set spending limits and merchant category restrictions — the kind of guardrails any bank would require before letting software spend money on a customer’s behalf. Mastercard’s Agent Pay specification requires explicit user authorisation for each transaction class. The pilot used cinema tickets and hotel bookings, which are exactly the low-risk categories you’d test first.
Is this going to scale, or is it a publicity demo? Both. The technology clearly works — two transactions settled cleanly through real-world payment rails. Whether it scales to the volume and merchant coverage of card payments is the open question. The friction is in the authorisation, identity verification, and merchant integration layers, not the AI decision-making. Adobe’s research on AI shopping behaviour suggests consumers are open to agent-led purchases for routine categories, but want strong controls.
🔍 THE BOTTOM LINE
New Zealand’s first agentic transactions went through Westpac NZ, IPSI, Mastercard, and an Australian LLM called Matilda. The pilot worked. The architecture is sound. The sovereign-AI question is open.
New Zealand has the talent, the strategy papers, and the demand. What it doesn’t yet have is a sovereign LLM with financial-grade deployment experience. Maincode beat NZ to it in February 2026. The question is who builds the next one, and where.
