Digital payments 2030 debate turns to small language models

Digital payments 2030 will be shaped less by faster infrastructure and more by AI systems embedded in transaction flows, according to Neeraj Aggarwal, senior project manager at Infosys Limited, in an external opinion published by Finextra. Aggarwal argues that the next competitive test for banks and fintechs will be how well their systems assess fraud, disputes, merchant onboarding, compliance and customer context in real time.

The view marks a shift from the past two decades of payments investment, which Aggarwal says focused on rails, cloud-native cores, API standardisation, tokenisation and cross-border interoperability. In his assessment, those upgrades improved speed and connectivity, but did not fully address the need for systems that can understand behaviour, intent and risk as transactions move through the network.

Finextra identifies the piece as external content from an author and says it expresses the author’s views and opinions. Aggarwal is listed by Finextra as a senior project manager at Infosys Limited.

What will define digital payments by 2030?

Aggarwal’s central claim is that small language models, or SLMs, will sit closer to the core of payments operations than large language models, or LLMs. An SLM is a narrower AI model built for a specific workflow, while an LLM is a broader model designed to handle many language tasks across domains.

He says payments firms need models that can respond within the tight timeframes required by real-time processing, operate inside a bank’s own environment and produce outputs that regulators and risk teams can audit. According to Aggarwal, LLM inference can take 300 to 800 milliseconds, while SLMs can run in 10 to 40 milliseconds, a gap he presents as material for real-time payment rails.

Data control is another part of the argument. Aggarwal says LLM APIs can require sensitive information to be sent to external processors, while SLMs can run on-premise in a way he says supports PCI-DSS, GDPR and emerging AI-risk rules. He also says SLMs can offer deterministic and auditable outputs for decisions tied to fraud alerts, anti-money-laundering reviews and disputes.

Where Aggarwal expects SLMs to be used

The payments use cases identified by Aggarwal include fraud detection, dispute handling, merchant onboarding, operations and customer experience. He says SLMs trained on institution-specific transaction patterns could detect anomalies, identify synthetic identities, score merchant risk and adapt to new fraud patterns.

For disputes and chargebacks, Aggarwal says SLMs could classify case types, summarise merchant evidence, identify missing documents, prepare policy-aligned narratives and recommend outcomes with audit trails. In merchant onboarding, he points to document extraction, know-your-business risk scoring, adverse media summaries and compliance classification as areas where automation could shorten approval times while highlighting riskier merchants earlier.

He also expects payment operations teams to use SLMs for exception management, reconciliation narratives, settlement discrepancy analysis and operational reporting. In customer experience, Aggarwal says LLMs will still have a role in conversational banking, personalised guidance, multi-document reasoning and customer-facing interactions, while SLMs handle controls closer to the transaction core.

How regulation fits into the payments AI argument

Aggarwal says global regulators are signalling that AI used in payments must be explainable, auditable, deterministic, governed and controlled for risk. He links SLM-based systems to requirements and frameworks including SR 11-7, DORA, Basel III, PCI-DSS and GDPR.

His proposed architecture for 2030 is a hybrid AI stack: LLMs at the edge for interaction and analysis, and SLMs at the core for fraud, compliance, disputes, onboarding, operations and real-time decisioning. For banks, fintechs and payment processors, the argument is that AI governance and reliability layers will become part of core payments infrastructure, rather than separate technology projects.

This story draws on original reporting from Finextra Research.