Ant International’s new payments model proves the future of financial AI isn’t just generative … it’s transactional

We have spent the last few years obsessing over artificial intelligence that sits in front of the customer – chatbots that summarise documents, copilots that draft emails, and interfaces that appear to know everything. But finance does not run on language. It runs on transactions, cash flows, liquidity, risk and trust.

After spending time examining the infrastructure powering global cross-border payments at Ant International’s recent Voyage event in Shanghai, I came away with a different perspective on where this technology is actually heading.

Ant International – which connects more than 150 million merchants and 50 digital wallets and banking apps through solutions like Alipay+ –  is proving that the next generation of AI will not be bolted onto financial products. It will be embedded directly inside the machinery of money itself.

That distinction becomes particularly important when we consider where commerce is heading. By 2030, AI agents are projected to manage somewhere between $3 trillion and $5 trillion of global consumer commerce, while the number of digital-wallet users is expected to exceed six billion. We are moving rapidly from a world where artificial intelligence tells us what we might want to buy towards one where artificial intelligence finds it, compares it, chooses it and, with our authority, pays for it.

At that point, the chatbot becomes one of the least interesting parts of the architecture.

Behind an apparently effortless transaction lies an extraordinary series of financial decisions about which account should be used, which currency, what FX exposure is being created, how much liquidity is required, whether the transaction is unusual, whether the agent is legitimate, who authorised it and whether the eventual transaction matches the intention of the human who gave the agent its instructions.

That is where the AI story collides with the financial infrastructure story.

The first generation of generative AI has largely been about putting intelligence in front of the customer. The next generation will increasingly put intelligence underneath the customer, embedded into treasury, payments, risk, identity, liquidity and transaction processing.

There is also a fundamental difference between applying AI to finance and using it to write an email or summarise a document: finance is an unforgiving environment for artificial intelligence. A chatbot telling you some obscure historical fact that is false is irritating. A financial model recreating  your company’s cash position, currency exposure or payment instruction is expensive.

The closer AI gets to the movement of money, the less tolerance there is for being approximately right. That, for me, is the real reason specialised financial models matter.

If AI is going to move from advising us about money to making and executing financial decisions, the benchmark changes dramatically. It is no longer simply whether the model can reason convincingly, but whether it can predict cash flows, recognise payment patterns, understand financial context and support autonomous decisions involving real money at enormous scale and with extraordinary reliability.

Ant International provides an interesting window into how this new financial AI stack is developing because it is approaching the problem from several directions at once.

There is:

• prediction, understanding what money is likely to do next;
• payments, understanding the transactions through which money moves;
• agentic commerce, where machines increasingly initiate those transactions themselves; and
• trust, where the financial system must determine whether those machines are who they claim to be and are doing what their human owners authorised them to do.

Seen this way, Ant International’s strategy is more interesting than a collection of AI products. It is an attempt to move specialised artificial intelligence from the edges of financial services towards its core infrastructure, and the newest evidence is not a chatbot but a payment model built to understand the transaction itself.

Specifically, at Voyage, Ant International unveiled the Antom 3-in-1 Transformer Model, which it describes as the world’s first 3-in-1 payment foundation model and its largest payment foundation model to date. It goes straight to the heart of what happens when AI moves inside the machinery of money because it is teaching AI to understand the payment.

The Antom 3-in-1 Transformer Model is designed around three forms of financial information at once:

• sequence data drawn from decades of payment records,
• graph data that maps relationships and behaviour, and
• tabular data rooted in risk-management expertise.

Ant International says the model has more than 10 billion parameters and is trained on 90 trillion tokens of real-world data annually. More strikingly, one unified model consolidates more than 200 legacy point solutions, turning what used to be a fragmented collection of specialist models into a common intelligence layer for payments.

The point is not scale for its own sake.

Payments demand decisions in milliseconds, and Ant International says Antom processes transactions end to end in around 25 milliseconds while covering payment protection, abuse protection and account security.

In its reported business results, the model delivers an average five-percentage-point improvement in payment success rates. This matters because it shows what a payment foundation model is supposed to do: understand enough of the transaction, the account and the surrounding behaviour to improve acceptance while reducing the cost of fraud.

This also changes the way we should think about AI architecture in finance.

Much of the industry has built hundreds of narrow models, each trained to solve one problem such as chargebacks, scams, suspicious trading, free-trial abuse or refund fraud. Ant International brings those problems into one foundation model, creating a system that is designed to anticipate threats as well as react to them.

In other words, the model is not being asked to talk about a payment after it happened. It is being asked to understand the payment while it is happening.

Antom Autopilot then takes the idea a stage further. Built on the same AI-native platform, it applies agentic AI across the merchant payment lifecycle, from onboarding and integration through payment optimisation, fraud protection, dispute recovery and global expansion. This is where the claim of a full-stack AI-native business starts to make more sense because the payment model is not isolated from the rest of the financial operation. It sits alongside account, FX treasury and growth capabilities, allowing intelligence developed in one part of the stack to support decisions elsewhere.

Then there is Ant’s Falcon Time-Series Transformer or FalconTST for short.

Antom 3-in-1 Transformer Model is about understanding the payment that is happening now, while the FalconTST Model is about understanding where the money is going next. Put the two together and the specialised-AI argument becomes clearer: finance needs models that understand both the transaction and time.

That is the problem Ant International set out to solve with FalconTST. Rather than building another model designed to understand words, FalconTST is designed to understand numbers changing through time and, critically, to predict what those numbers will do next. In finance, that means forecasting cash flows, liquidity requirements and foreign-exchange exposures, turning the abstract idea of specialised financial AI into something that has a direct impact on how businesses manage their money.

Banks and corporations live in a world of constantly changing numbers.

Money arrives and leaves, currency exposures rise and fall, customers behave differently according to the hour, day, season and economic environment, while markets, liquidity and risk move continuously. For a multinational company processing millions of transactions across currencies and markets, understanding those patterns is not an academic exercise because getting them wrong costs money.

Consider an airline selling tickets around the world. It might receive dollars, euros, yen, baht and dozens of other currencies while paying for aircraft leases, fuel, airport charges, staff and suppliers across an entirely different pattern of currencies. Somewhere inside that enormous flow the company has to work out how much money it will receive, where it will receive it, when it will arrive and therefore what currency exposure it needs to hedge.

Forecast too much and it over-hedges, while forecast too little and it remains exposed. Multiply those decisions across currencies, countries, businesses and millions of transactions and tiny forecasting errors turn into serious amounts of money.

FalconTST was built for precisely this world. Rather than concentrating primarily on words, as an LLM does, it concentrates on numbers changing through time and is designed to identify cycles, trends, seasonality and sudden changes across time-series data, creating a forecasting foundation model that can be applied across financial and commercial environments.

That difference matters because finance is fundamentally temporal. A balance means something at a particular moment. A currency exposure exists over a particular period. Cash flow has a rhythm, while markets have patterns, interruptions and shocks. Understanding finance therefore requires more than understanding what a number means. It requires understanding what that number is likely to do next.

Of course, finance has a rather brutal benchmark for all of this: did it save money?

This is where FalconTST moves from being an interesting AI research project into something far more significant. The technology originated inside Ant International, forecasting cash flows and foreign-exchange exposures across hourly, daily and weekly periods, and is now being incorporated into financial solutions involving banks including Barclays, Citi, Deutsche Bank and Standard Chartered.

The practical example I find most interesting is Capital A, whose businesses include AirAsia. The company used Ant International’s time-series technology to forecast sales and foreign-exchange exposure across its multi-currency operations. According to Ant International, forecasting accuracy exceeded 90% and FX hedging costs fell by around 40%.

That is where the debate about financial AI gets real. A model that writes a better email creates productivity. A model that predicts tomorrow’s cash position more accurately means a company can manage liquidity more efficiently, hedge currency exposures more precisely and put less capital aside to protect itself against uncertainty.

That is the moment when AI stops being a PowerPoint strategy and starts becoming financial infrastructure. Antom shows specialised AI learning to understand, protect and optimise the payment in real time, while FalconTST shows specialised AI learning to forecast the cash flows and exposures that follow.

The wider point from Voyage is that the future of financial AI will not be defined by one giant model trying to do everything. It will be defined by specialised intelligence embedded across the stack, connecting payments, accounts, treasury and growth so that money can be understood before it moves, while it moves and after it moves.