Payments industry experts highlight the importance of specialised, ultra-fast models for fraud detection and proposed a two-lane approach that balances speed with human oversight, challenging the widespread use of large language models in high-speed transaction environments.
In a payments business, the temptation to put a large language model everywhere is strong. But as Shashank, who runs technology at a payments company, argues, the authorization path is the wrong place for a slow, chatty model. In his account, card transactions move through multiple systems and leave only a narrow timing window for fraud checks, making a 400 millisecond model call unsuitable when a terminal is expecting a near-instant answer. That constraint helps explain why many real-time fraud vendors focus on millisecond-level scoring rather than generative models. Argus Mesh, FraudSpark, Kapix, Stripe AI, InferX and Stealth IQ all describe their systems in terms of sub-50 millisecond or even sub-10 millisecond decisioning, which reflects the speed demands of payments infrastructure.
The alternative Shashank describes is a two-lane design. An offline machine-learning model, in his case a gradient-boosted decision tree, makes the first-pass decision in microseconds, while uncertain cases are pushed to human review. He says only a small share of traffic falls into that grey zone, where an LLM can be useful not as judge and jury but as a summariser. That division of labour mirrors the way fraud platforms market themselves: vendors such as Argus Mesh and FraudSpark emphasise real-time scoring and decision workbenches, while InferX and Kapix pitch continuous signal updates and automated risk assessment for production environments.
Shashank says he tested the approach on the IEEE-CIS fraud dataset, training an XGBoost model on 472,000 rows and validating it on 118,000 more. He reports an AUC of 0.9436, which he presents as strong enough for production use, and says the full scoring path, including explanations, took about 6 milliseconds on a laptop. He also argues that the model surfaced the sort of velocity-based signals payments engineers already trust, reinforcing the idea that narrow, well-understood models can outperform broad AI claims in the part of the stack where speed matters most.
The deeper point is not that AI has no place in fraud detection. It is that the tool has to fit the task. Shashank’s position is that LLMs are best used where humans need fast comprehension, not where systems need immediate decisions. In his example, the model writes a concise case summary for analysts, turning a string of raw fields into something readable in seconds. That framing also serves as a warning to fintechs still selling LLMs as a shortcut for real-time fraud control: if the latency profile is wrong, the architecture is wrong too.
Disclaimer: This article is intended to inform and educate, not to recommend or endorse any financial product, investment or strategy. Please consider your own financial circumstances and seek professional advice where appropriate before making financial decisions.
