Page 34 - gs260802

Insights and Expertise

How AI and LLMs will Why LLMs change the equation

The emergence of large language models (LLMs) as a prac-
transform payment tical business tool has introduced a capability that earlier
AI architectures could not easily provide: the ability to
gateway technology interpret complex, high-dimensional data and surface it
in natural language that non-technical users can act on
immediately. This is the distinction that matters most for
payments.

Traditional machine learning (ML) models have long been
applied to fraud scoring and transaction routing, and
they've delivered genuine value. Mastercard's Decision
Intelligence Pro, built on a generative AI recurrent neural
network, can scan 1 trillion data points and has improved
fraud detection rates by an average of 20 percent, reach-
ing as high as 300 percent in certain cases, while reducing
false positives by more than 85 percent (see https://tinyurl.
com/7bdkese).

A separate Mastercard generative AI system focused on
compromised card detection doubled its identification
rate and increased the speed of flagging at-risk merchants
by 300 percent (see https://tinyurl.com/4mfdww3r). These are
the kinds of operational outcomes that drive boardroom
conviction.
By Andrii Shevchuk
CONCRYT But pure ML models, however powerful, are opaque. They
produce scores and signals; they do not explain them-
or years, artificial intelligence (AI) in payments selves in language that a merchant's operations team can
was treated as a horizon technology, something digest and relay to their finance director or customer ser-
to watch rather than deploy. That era is now vice staff. And LLMs change this fundamentally.
F over. According to a major 2025 industry report
by HCL Tech (see https://tinyurl.com/ysbsw2md), 99 percent By sitting on top of transaction data, risk signals and de-
of organizations are now using AI in payment operations, cline codes, an LLM layer can translate raw outputs into
and 52 percent expect to become fully autonomous finan- actionable plain-language explanations. A gateway aug-
cial service providers within the next 18 to 24 months. mented with LLM capability can tell a merchant not just
that a transaction failed but why, which bank issued the
These colossal figures reflect a structural shift already un- decline, whether the pattern is recurring, and what reme-
derway across the global payments landscape, one that's diation is available. This closes the intelligence gap be-
rewriting the expectations merchants, acquirers, and pro- tween data and decision.
cessors have of the technology sitting between them and
their customers. Stripe, which unveiled what it described as the world's
first AI foundation model for payments at its annual Ses-
The payment gateway, long regarded as a reliable but un- sions event in May 2025, has moved particularly aggres-
glamorous piece of infrastructure, is at the centre of this sively in this space. The company built its agent toolkit to
transformation. Gateways have historically served a nar- allow LLMs to interact directly with payment functions
row function: route the transaction, return an authoriza- through natural language calls, integrating with OpenAI's
tion code, pass the result back. But this model was built for Agents SDK, LangChain and other major frameworks.
a simpler era.
Stripe's developer documentation explicitly frames LLM
Today's payments environment demands far more. Mer- integration as the new architecture for intelligent payment
chants need real-time intelligence on why transactions are workflows, enabling capabilities that range from dynamic
declining, how their acceptance rates compare across mar- routing to automated financial reconciliation.
kets, where fraud risk is concentrating, and what they can
do about any of it before the next payment attempt arrives. The competitive advantage is real, and it's widening

The gateway that merely processes is being replaced by For merchants operating at scale, the business case for
the gateway that thinks. AI-enriched payment gateways has moved well beyond
theory. Acceptance rate optimization is perhaps the most

34