Financial institutions have largely used generative artificial intelligence (AI) to summarise documents, generate code, answer customer queries and automate administrative work. These applications improve productivity but preserve the division of responsibility between software and people. AI assists, while people decide and act.

Timothy Wong, Global Vice President of Data and AI at Airwallex, expects that boundary to shift. “In an AI world, we will try to understand the intent of the user, provide recommendations, and then sometimes just complete the action on behalf of users if it’s low risk,” he said.

Users could specify an outcome instead of navigating a sequence of screens and approvals. An AI system would assemble information, recommend an action and, within defined limits, execute it. “The UI part has actually become surface agnostic as well,” Wong said. The same financial capability could be accessed through different applications, platforms or AI interfaces.

That shift is drawing payments and technology companies into a broader effort to build the infrastructure for agent-led commerce and financial services. Stripe has developed an agentic commerce suite, Visa and Mastercard are building credentials and controls for agent-led payments, and Google has introduced protocols for agent communication, payment authorisation and commerce.

Airwallex’s case is that AI models will become increasingly accessible, while licences, payment connectivity and controls for executing financial transactions will remain harder to replicate.

Airwallex sees payment infrastructure as its AI moat

Airwallex is a global payments and financial platform for businesses. Its core services allow companies to collect payments, hold and convert multiple currencies and pay suppliers across borders, often using local accounts and domestic payment rails instead of routing every transaction through correspondent banking networks.

The company raised $320 million in June 2026 at an $11 billion valuation. In March 2026, it reported $1.3 billion in annualised revenue, $287 billion in annualised transaction volume and more than 676,000 businesses served directly or through platform customers.

According to Airwallex, more than 90% of its transactions are routed through local payment rails. It has accumulated 89 licences and permits globally, and direct integrations into more than 160 local payment methods, with payouts across over 200 countries and territories. Wong described those licences as the product of more than a decade of compounding investment.

A general-purpose AI model may interpret an instruction to convert currency, issue a card or pay a beneficiary. It cannot complete the transaction without authenticated accounts, customer data, payment connectivity and rules governing what the user is permitted to do.

That distinction is central to Airwallex’s positioning. AI models can be sourced from several providers, but licences, global payment connectivity and the ability to execute transactions through controlled APIs are harder to replicate. “The moat actually is not being replaced because of AI,” Wong said. “Ideas are cheap,” he added. “But in order to be really successful, first you need to have a moat.”

Airwallex does not own this opportunity. Banks hold regulated customer relationships, deposits and transaction data. Payment networks possess global acceptance and credential infrastructure, while enterprise software providers control corporate systems of record.

Wong said infrastructure alone was insufficient. He pointed to leadership willing to invest early in AI-native products and an engineering team able to develop and test them.

The company is also applying AI internally through what Wong called a “horizontal and vertical” model. OpenAI and Anthropic provide the general-purpose AI layer, while Airwallex builds its own tools and redesigns functions such as customer relationship management and marketing. “It’s not really a transformation anymore, it’s establishing this new status quo,” he said.

One example is AirForge, an in-house platform that allows non-technical employees to build and share AI applications. Engineers can refine the applications while Airwallex applies security and compliance controls centrally. Wong said these tailored systems could become an organisational differentiator.

Kai, T:0, AgentOS and Airi extend Airwallex’s AI strategy

Airwallex’s products can be viewed as different parts of an emerging AI-enabled financial architecture. Kai is its customer-facing assistant, T:0 is designed to unify enterprise finance records, AgentOS connects external agents to Airwallex’s capabilities and Airi is intended to provide payment credentials and controls for agent-led commerce. The products are at different stages of development.

Kai was introduced as Airwallex AI Assistant in February 2026 and renamed in June. It is embedded in the company’s web application and supports onboarding, customer queries and selected account operations. Airwallex said in June that Kai handled 88% of user conversations without human escalation and had helped more than 14,000 customers through onboarding.

Wong said Airwallex was working to connect Kai to more of its financial capabilities. One prospective use would combine information across accounts and recommend action when a business has excessive exposure to a particular currency. This would extend Kai beyond customer support into financial recommendations based on account data.

T:0, announced in private beta on 25 June, addresses the system-of-record layer. It is intended to automate bookkeeping, forecasting, tax, compliance and reporting as an AI-native finance platform, instead of adding an assistant to an existing enterprise resource planning system.

Finance teams typically work across accounting platforms, banking portals, treasury systems and spreadsheets. “We are trying to unify the system of records,” Wong said. T:0 is also being designed around specialised agents, including a CFO agent. Wong said Airwallex was examining the decisions finance professionals make, the information they use and the actions that follow, before standardising those processes through APIs.

AgentOS provides the connection layer for external AI agents. Launched in June, it combines a command-line interface, a Model Context Protocol server, plugins and predefined skills for activities such as creating beneficiaries, issuing cards, producing cash-flow reports and turning contracts into invoices.

Wong said AgentOS could let businesses connect agents without navigating as much API documentation or relying as heavily on Airwallex’s solution engineers. AgentOS is available, although its command-line interface and predefined skills remain in beta. By default, it does not execute money-out actions autonomously; payouts, currency conversions and transaction approvals require human confirmation.

Airi is positioned to address the payment layer. Airwallex launched it on 25 June as a one-click checkout product built into Airwallex Checkout and Payment Links. The company said it increased conversion by up to 14% in early testing.

Airwallex intends to develop Airi into an agentic wallet through which shoppers could authorise agents to transact on their behalf. Its roadmap includes delegated agent payments, spending limits, permission controls and multicurrency balances. These capabilities are not yet available. Wong described the planned product as an “agentic wallet concept”, arguing that existing wallets were designed primarily for people.

Airwallex links AI autonomy to financial consequence

For financial institutions, the central control question is which actions an AI system may complete autonomously, and which should continue to require human approval.

Airwallex classifies requests according to their nature and the consequences if they are handled incorrectly. An informational query, customer-support problem and instruction involving funds present different risks.

Airwallex uses AI and, in some cases, hard-coded logic to identify the type and intent of a request. It then routes the request according to its complexity and potential consequences. “The money movement itself, we still [execute through] our API, which is a deterministic action,” Wong said.

This separation allows the model to handle ambiguous language without giving it uncontrolled authority over funds. Airwallex uses online evaluation to monitor systems in operation and offline human review to calibrate performance. Wong said the methodology is reused across the company, but its application varies because a customer query, know-your-customer (KYC) decision and payment instruction carry different consequences.

Wong’s approach treats autonomy as a continuum, with the level of human involvement determined by the consequences of error. Many financial operations will continue to require explicit consent. Accountability also remains with the people and organisations deploying the system. “I think the human is still ultimately accountable,” Wong said. “I don’t think that we should blame AI.”

Trust is also a user-experience problem. Extensive checks do not create confidence if customers cannot see them, but presenting every control may overwhelm them. Airwallex uses A/B testing and customer reviews to determine how much assurance information Kai should display.

Agent protocols improve interoperability, but compliance gaps remain

Controls within one platform are insufficient when agents operate across organisational boundaries. Open standards are emerging to connect agents with other systems. The Model Context Protocol (MCP), developed by Anthropic and now governed under the Linux Foundation’s Agentic AI Foundation, connects AI applications with external tools and data. The Agent2Agent protocol, originally developed by Google and now hosted by the Linux Foundation, allows agents built on different platforms to discover, communicate and coordinate with one another.

At the commerce layer, Google’s Agent Payments Protocol (AP2) uses cryptographically signed mandates to record user intent and agreed transaction details. Its Universal Commerce Protocol provides common functions across product discovery, checkout and post-purchase services. Visa’s Trusted Agent Protocol allows merchants to verify an agent’s identity and associated authorisation, helping them distinguish legitimate agents acting for customers from unidentified automated traffic.

These protocols support communication, verification and authorisation, but do not themselves execute payments. Nor do they determine whether a transaction complies with an institution’s internal policies and regulatory obligations or resolve liability when an agent acts incorrectly. Financial institutions must still apply their own account permissions, regulatory controls and audit requirements before executing a transaction.

Airwallex intends to support several protocols instead of assuming one will dominate. “Maybe there will be one player, maybe not,” Wong said. “It’s still to be seen.”

If AI agents become a primary interface for commerce and financial services, competitive advantage may increasingly belong to companies that can execute authorised actions safely and at scale.

Airwallex has spent more than a decade building licences, local payment connectivity and APIs for global money movement. The question is whether that infrastructure, combined with its ability to develop and test products quickly, gives it an enduring advantage as financial activity shifts from people navigating software to agents acting on their intent.