Page 24 - gs260102
P. 24
Cover Story continued
Fraudsters will open a bank account at a fintech or nano
bank, Kopelman explained, then put a fake business name
Agentic commerce risk: on the bank's letter.
Where it breaks, how to contain it
He's also heard underwriters raise concerns about phony
As agentic commerce moves from experimentation applications and bust outs, in which fraudsters run
to production, payments leaders have said the core massive numbers of stolen credit cards as soon as their
challenge is not innovation speed but risk contain- merchant account is activated. Then they disappear.
ment. History suggests that new payment capa-
bilities rarely fail all at once; instead, weaknesses LLM hallucinations
emerge at specific points in the transaction lifecy- Thomas Mueller, co-founder and CEO at Rivero, a European
cle. Industry experts have identified several recur- fintech, advocated the use of rules-based AI in commerce
ring risk zones that financial institutions and com- and banking applications, stating that these technologies
merce platforms must address as AI agents begin deliver a more consistent and reliable customer experience
acting on behalf of consumers and businesses. than large language models (LLMs), which can be prone to
hallucinating and making things up.
1. Identity and authorization risk
"There's a lot of excitement about AI, but people who see
At the foundation of agentic commerce lies beyond the hype recognize the need to map problems to
the question of who, or what, is acting. Un- solutions in a secure, predictable way," he said. "I would
like traditional transactions initiated by a urge banks and service companies not to expose customers
human user, agentic systems introduce non- to a large language model or work with a fintech that uses
human identities that can make autonomous LLM for customer service."
decisions.
Air Canada learned this the hard way, Mueller noted,
Risk arises when agents operate without clear when a chat bot made up a refund rule that didn't exist.
identity, authority scope, or revocation mech- When the consumer went to claim that refund, the airline
anisms. Experts recommend cryptographic was held liable for the chatbot's actions and had to refund
authentication, credential tokenization and the flight. This highlights why LLMs are not ideal for
agent registration frameworks similar to building virtual agents and apps, he said.
those used for issuing payment credentials.
Emerging standards such as Visa Trusted Regulators are also taking a hard look at AI, Mueller
Agent Protocol and Mastercard Agent Pay are pointed out, citing the European Union's Artificial
designed to help distinguish trusted agents Intelligence Act, enacted in March 2024 and designed to
from malicious bots. be phased into law over a two-year period.
2. Decision drift and hallucination The regulatory framework makes it more difficult for
banks and B2B fintechs to build products on top of machine
Large language models (LLMs) can introduce learning models that are not deterministic or explainable,
unpredictability into commerce flows by he said, adding that he agrees with and supports these
generating incorrect or fabricated responses, protections.
a phenomenon commonly referred to as hal-
lucination. Multilayered protections
Several industry leaders caution against de- Troy Leach, chief strategy officer at Cloud Security
ploying LLMs directly in customer-facing Alliance, a not-for-profit organization specializing in
payment or authorization workflows. In- cloud computing security best practices, highlighted the
stead, they advocate rules-based or deter- need to balance agentic AI's speed and convenience with a
ministic AI for transactional decisions, with proportionately robust security framework.
clearly defined logic paths and constraints.
The goal is not to eliminate AI reasoning, but Leach said the key is "to understand the risk and limitations
to prevent agents from improvising in regu- associated with the technology and to understand this
lated or high-risk scenarios. requires a layered approach, not only because the security
is still being defined as AI advances but because we have
3. Fraud amplification and scale taken the most complex part of the transaction equation,
the consumer, and added the complexity of a non-human
Agentic systems increase the speed and scale that could drift into non-deterministic decisions."
at which fraud can occur. Automated agents
Agents that act on behalf of humans need to be carefully
designed, Leach noted, to prevent legal violations, such as
24
