Who Pays When AI Forecasting in Finance Gets It Wrong

The question put to Kelvin Li was a broad one: where financial AI is heading once forecasts this accurate stop being exceptional. Li, General Manager of Platform Tech and Senior Vice President at Ant International, answered with the terms on which his company gets paid.

“When the model accurately delivers a forecast result, we will be commercially rewarded. When it does not, we then take less revenue, or even no revenue,” he told The Bright Minded. The model is FalconTST, which Ant built to forecast cash flow, liquidity and currency positions, and Li described the arrangement in terms of its work for banks.

Ant announced the latest version of FalconTST in August with a benchmark result, a state-of-the-art score it reported on a common measure of forecasting error. Li’s answer adds a second test for AI forecasting in finance: what the provider loses when the number turns out wrong.

A wrong forecast already has an owner

A currency forecast tells a business how much of each currency it will hold, and when. A business that hedges more currency than arrives has paid to protect money it never held, and one that hedges less leaves the gap exposed to exchange-rate moves. Banks, corporate treasuries and financial technology providers all make these calls.

The market those hedges pass through is vast. Trading in foreign exchange (FX) averaged $9.6 trillion a day in April 2025, according to the latest Triennial Survey by the Bank for International Settlements (BIS). Outright forwards, the contracts used to lock in a future exchange rate, accounted for $1.8 trillion of that, or 19% of the total, against 15% in 2022.

In the UK, responsibility for a bank’s models already has a clear owner. The Prudential Regulation Authority’s model risk principles cover models built in-house or bought from vendors, regardless of the technology behind them, and allocate responsibility for the framework to a senior manager. They apply to banks with approval to use internal models for their capital calculations.

The gap Ant says general AI leaves open

Li’s case for a specialist model starts with what general AI cannot promise. “Our big bet is that Large Language Models alone are not enough to solve problems in the financial industry,” he said. Large Language Models (LLMs), the systems behind chatbots, work with language, and Ant’s answer was a Large Data Model (LDM), built to predict the next data, which Li says helps the company manage its risks and price its financial services.

The deeper problem, in his view, is how those general systems reach an answer. “AI for finance will require decisive technology for a highly regulated industry, whereas current AI technologies such as LLMs are probability-based,” Li said. The pricing arrangement is what he presented as one way Ant closes that gap.

Payments carry a version of the same tension. Contributors to The Bright Minded argued, in a piece on the deterministic advantage fintechs already hold, that AI agents decide on probabilities while payment systems move money on instructions they assume are correct.

The model also answers a more practical gap. Ant built FalconTST because it saw that “there are still certain gaps in traditional banking services that cannot meet the needs of a digital economy,” Li said, naming fixed banking hours and bank holidays among them. Ant reports that the model has reached over 93% accuracy in FX and cash-flow forecasting.

What four banks let Ant see

Barclays, Citi, Deutsche Bank and Standard Chartered have built FalconTST into their own FX systems, as The Bright Minded reported when Ant released FalconTST 2.0 in August. At Standard Chartered, the model works inside PathFin.ai, a programme run by the Monetary Authority of Singapore, the city-state’s central bank and financial regulator.

Those partnerships changed what Ant could observe. “AI for finance is still in the early stages of development,” Li said, adding that working with banks gave Ant “a glimpse of how AI will interact with core banking systems”. Until then, he explained, Ant had worked only with banks’ supporting and operating functions.

Li argues that Ant also reads something the banks do not. He calls FalconTST a “user-generated solution”, developed from problems Ant and its customers met in their own operations. Banks, he said, are “not as closely connected to the underlying commercial scenarios behind their customers’ transaction data, whereas FalconTST analyses the commercial activities underlying a business’s payments.”

A price that moves with the error

The pay by performance model, as Li calls it, follows the bank’s own economics. “We align ourselves with the bank’s economic value, as the model’s forecast accuracy decides the bank’s risk and cost,” he said. An accurate forecast lowers the bank’s costs and raises Ant’s share of revenue, so the bank and its provider gain or lose on the same outcome.

Paying AI providers for results appears outside finance too. Intercom, which sells an AI customer service agent called Fin, prices it per outcome from $0.99, with no seat licences required.

Li describes Ant’s longer-term role in terms of banks, businesses and regulators. “We see ourselves as the bridge between banks and businesses, working closely with regulators to support the scalable adoption of financial AI technologies,” he said.

Finance extends its trust to the party that loses money when the numbers are wrong, and a forecast priced on its accuracy puts the AI provider among those parties.

Editor’s note

Every piece published on The Bright Minded goes through careful verification, but mistakes can happen. If you spot an error, have additional information, or want to flag anything, write to rosalia@thebrightminded.com.