Airwallex: Scale local relevance with accountable AI

A single platform interface is efficient to maintain, but a business customer still operates within local currencies, payment practices and eligibility rules. Translation cannot fix a missing payment method, unexpected settlement treatment or an unavailable service. The strategic task is to preserve a coherent core while adapting the elements that materially affect a customer's ability to complete a financial task.

A four-layer framework makes these dependencies visible. Language improves comprehension; transactional localisation supports familiar payment choices; infrastructural localisation connects domestic financial systems; institutional localisation determines what may be offered. The layers interact, while serving distinct functions. Cultural and behavioural preferences matter alongside the economic and institutional conditions that govern a financial service.

Airwallex's Hosted Payment Page documentation describes supported languages, browser-locale behaviour and filters for available payment methods. Its FPX documentation provides a concrete Malaysian example with eligibility, currency and settlement conditions. This primary evidence shows adaptation at the transaction level. It also demonstrates why a generic claim of local availability needs qualification by entity, account and payment configuration.

The proposed diagnostic would follow a merchant and shopper journey from eligibility checks through checkout, settlement, reconciliation and exception handling. Where does a customer encounter an unexpected currency, unfamiliar wording or unexplained restriction? National averages can inform research priorities; they should not determine an individual customer's preferred language, payment method or tolerance for automation.

Airwallex's December 2023 announcement describes an LLM-based website scanner that distinguishes risky keywords from their business context. It reports fewer false-positive alerts and a 20% increase in customers moving through fully automated KYC, relative to its earlier process. The evidence points to reduced unnecessary review in a defined workflow; the absolute share of customers automated is not established by that figure.

Those results are company-reported and concern a component of due diligence. They do not prove lower financial crime risk or accurate performance in every language. The recommended role for AI is bounded: assemble relevant context, produce a reasoned explanation and route uncertainty to authorised reviewers. Customer-facing guidance should use approved product and jurisdiction information, with actions governed by existing permissions.

A global minimum product keeps costs low but can create avoidable local failures. Separate products for every market improve tailoring but multiply maintenance and control requirements. A modular model offers a middle course: shared identity and workflow foundations with configurable language, payment methods, eligibility and support content, owned jointly by platform and market teams.

The AI investment should follow the same discipline. Unrestricted conversational execution creates a wide operational surface before value is established. A read-only explanation pilot is easier to inspect and reverse. More consequential tasks would require separate validation, confirmation and authorisation. The IMF's financial-sector risk note reinforces the need to evaluate hallucination, bias, privacy, robustness and cybersecurity alongside convenience.

Define a baseline for eligible onboarding completion, time to a decision, repeat document requests and support contacts. Assess false positives against reviewer-adjudicated examples and evaluate false negatives with an appropriately governed test set. Review results by language, market and business type; a lower average review workload can conceal a poorly served subgroup.

For the customer pilot, test task completion and explanation accuracy using realistic, permissioned scenarios. Measure whether customers understand fees, settlement and escalation options. Finance should track support cost and contribution from retained, active customers after localisation costs. Expansion requires a credible gain without weakened risk outcomes or misleading guidance; faster completion alone is insufficient evidence for scale.