Agentic AI Moves From Experimentation Into Core Payments Operations

Agentic artificial intelligence is beginning to move from experimental banking applications into the operational heart of payments, as financial institutions look for ways to manage increasingly fragmented infrastructure with greater speed and resilience.

Agentic AI Moves From Experimentation

A new Finextra impact study produced with Volante argues that AI can act as an intelligence layer across existing payment rails rather than replacing them. Banks are already using AI alongside systems such as Faster Payments, BACS and CHAPS to analyse transaction data and determine whether payments should proceed, be challenged, rerouted, investigated or enriched.

The potential gains are significant. Volante estimates that 30% to 40% of payments still require some form of manual intervention, leaving considerable scope for automation.

AI Expands Beyond Fraud Detection

Fraud prevention remains one of the most established applications, with the report estimating that around 90% of financial institutions now use AI-powered fraud solutions.

But agentic AI is expanding into more complex areas.

Payment routing systems can dynamically choose between processors, gateways and acquiring banks according to approval rates, cost and settlement speed. AI can also enrich transaction data, support behavioural authentication and improve intraday liquidity management.

The report envisages increasingly autonomous systems capable of responding to cash positions by funding accounts or rerouting payments without waiting for manual instructions.

Through intelligent enrichment and adaptive routing, Volante says some users have achieved straight-through processing rates of up to 95%, while reducing exception-handling costs.

Autonomy Must Increase Gradually

The report nevertheless cautions against giving AI unrestricted control over payment processes.

Its recommended model is “progressive autonomy”, beginning with assisted decision-making and increasing an agent’s authority only as performance evidence demonstrates reliability.

Four categories of agent are proposed: ‘prevent’, designed to stop failures before they occur; ‘repair’, which resolves problems automatically; ‘predict’, which identifies the best outcome for individual payments; and ‘sense’, which continuously monitors risks and operational congestion.

Higher-risk decisions require correspondingly stronger human involvement. Customer-service summarisation could potentially operate autonomously, while payment releases, sanctions decisions and account restrictions should retain human decision-makers.

Governance Becomes Part of Payments Architecture

For banks, explainability is particularly important. AI agents should provide decision rationales, identify their underlying data sources, assign confidence scores and maintain complete records of prompts, outputs and actions. Historical decisions must also be capable of being reconstructed for investigators, auditors and regulators.

The report advocates human review queues, immutable audit logs, transaction thresholds, dual authorisation for higher-risk actions and emergency “kill switches”.

Business continuity also matters. Banks need fallback procedures allowing payment processing to continue when AI services fail or produce unexpected outputs.

The significance of agentic AI is therefore not simply that it can automate payments. Its longer-term impact could be to create an intelligent orchestration layer across fragmented banking infrastructure — provided autonomy grows alongside governance, explainability and human accountability.