Airwallex: AI Redefines Finance with Initial Challenges

Deploying AI in the Financial Sector: First Steps and Pitfalls to Avoid

Key Points

To effectively integrate artificial intelligence (AI) into the financial sector, it is crucial to start with one or two high-impact use cases. These cases should allow AI to have a tangible effect within the first few weeks, relieving teams from repetitive tasks. Decisions regarding infrastructure are more critical than merely listing AI features. Disconnected systems can limit AI capabilities.

Airwallex, for example, brings together a global financial infrastructure and software on a single platform, providing AI with complete visibility over banking data, payment acceptance, and expense management. The sudden omnipresence of AI hints at possibilities where it could manage our accounts, predict cash flow deficits, and automate financial operations. However, these promises can be confusing, and it is essential to know where to start.

Successful teams begin on a small scale, focusing on work efficiency. They wait for the foundations to operate reliably before scaling up. This approach allows for transforming grand promises into steady progress and ambition into concrete actions.

Areas Where AI Quickly Adds Value to Finance Teams

AI has already proven its effectiveness in certain aspects of finance. High-volume processes with clearly defined outcomes particularly benefit from AI. For instance, managing employee expenses is a time-consuming process, as each receipt must be verified and categorized. AI accelerates this process through automatic data extraction and expense categorization.

Payment failures disrupt cash flow and require manual follow-ups. AI improves payment success rates by identifying the most effective payment routes and retry strategies for each type of transaction. Airwallex uses AI to enhance payment acceptance rates and reduce false positives.

In terms of anti-fraud controls, traditional rule-based systems often block legitimate payments. AI models assess risk using comprehensive transaction history, improving reliability and reducing unjustified alerts. Mastercard reports that its Decision Intelligence Pro system has improved fraud detection by an average of 20%, and up to 300% for certain deployments.

Account reconciliation is another task where AI can make a difference. It continuously compares transactions and flags issues at an early stage. A study from Stanford University and MIT found that accountants using generative AI closed monthly accounts 7.5 days earlier and freed up about 3.5 hours per week for higher-value tasks.

How to Deploy AI in the Financial Sector: A Practical Guide

Step One: Choose Two Defined Use Cases

Most teams achieve better results by starting with targeted and measurable actions. Choose a process that slows down your team and one where errors and delays are unforgiving. For most companies, this typically involves business expenses, anti-fraud checks, or payment retries. The goal is to demonstrate that AI can reduce your workload within a reasonable timeframe.

Step Two: Assess Your Infrastructure

AI performs better when it has access to your bank statements, cards, and financial data all in one place. If these elements are scattered across different systems, you may need to reconnect them through integrations or adopt a unified platform. In any case, the most important thing is to have a clear understanding of the costs and efforts involved to avoid surprises during the project. Starting with disconnected tools often reveals that integration costs are higher than migrating to a unified platform. Do your calculations carefully before deciding which path to take.

Step Three: Define Policies and Set Limits

First and foremost, establish basic principles: spending limits, approval rules, escalation procedures, and definitions of exceptions. The clearer your rules are, the more AI can assist you. Additionally, these rules will reassure your team that everything happening behind the scenes remains under control, traceable, and easy to reverse if necessary.

Step Four: Pilot with Human Oversight

In the initial weeks, allow AI to suggest actions that your team will review and validate. This practice gives you insight into the model's behavior, its added value, and points to monitor. Scrutinize both the gains (time saved, fewer rejections, better account reconciliations) and any potential errors. A weekly review should be sufficient to keep things moving forward.

Step Five: Expand Usage to Adjacent Processes

Once you see concrete results in one area, move on to the next process. If everything has worked for expenses, move on to billing. If fraud detection has improved, add payment optimization. The trick is to keep everything connected to avoid creating new silos. Growth should take the form of a system being built rather than a series of mini-experiments being pieced together.

“We are not just interested in incremental improvements. We are reimagining how businesses manage their finances. Across all our products, we systematically integrate artificial intelligence to eliminate manual processes, enable more informed decision-making, and provide true autonomy to our clients.” Shannon Scott, Chief Product Officer, Airwallex

Measuring the Success of Your AI Deployment

Often, teams either complicate evaluation by tracking dozens of unnecessary metrics or avoid it altogether because the notion of success seems subjective. A good compromise is to focus on 3 to 5 concrete metrics related to your use cases, establish a baseline before any intervention, and monitor both quantitative progress and qualitative signals relayed by your team as they use the system.

Choose Relevant Metrics for Your Use Cases

The metrics to track depend entirely on your starting point. In any case, take note of the baseline figures before changing anything. Otherwise, you will only be guessing whether things are improving or if it’s just an impression.

For example, if you decide to apply AI to one of the areas mentioned above, we suggest the following metrics for each type of process:

• Employee Expense Management: average processing time per expense report; exception rate requiring manual processing; and time between request and reimbursement.

• Payments and Payment Retry Attempts: acceptance rate distributed by site or card type; specific reasons for refusals; and frequency of successful payment retries.

• Fraud Control: false positive detection rate and actual fraud detection rate; as well as the average time your team spends reviewing each flagged transaction.

• Account Reconciliation and Closing: match rate, number of exceptions requiring manual verification, and number of days needed to close accounts compared to your current process.

Be Vigilant for Friction Points

You will know your AI model is working when your team stops devising workarounds or manually checking each result before proceeding. Conversely, a deficient AI model creates new forms of work instead of eliminating old tasks. Teams then spend their time correcting miscategorized expenses, reversing nonsensical routing decisions, or explaining to stakeholders why the system flagged a legitimate transaction as suspicious. Weekly check-ins during the pilot phase should reveal any patterns early enough to make adjustments before these pitfalls become habits that could sabotage the deployment.

Trust Qualitative Signals

The numbers tell you what has changed, but it is how your teams talk about AI that will reveal how it will be adopted in the long term. The best indicator of deployment success is when teams start proposing new use cases without being prompted. This means they are no longer skeptical and see AI as a tool to make their lives easier. Once trust is established, questions shift from “Can we trust this thing?” at the beginning of the pilot phase to “Could we also use it for billing processing?” Teams that trust the system stop conducting parallel manual processes.