From RPA to AI Agents: The Next Banking Revolution

How banks can evolve from deterministic automation to bounded autonomous operations

RPA Automated Execution

For more than a decade, banks transformed operations through robotic process automation. I was part of that journey, implementing automation across banking processes where speed, accuracy, scalability and control mattered.

Trade finance provides a powerful example. Consider letter-of-credit issuance. A customer submits an LC application. The process requires data extraction, field validation, rule checking, clause review, limit verification, LC creation, accounting entries, SWIFT preparation and transmission, downstream system updates, and customer notification.

In implementations I was involved with, mature automation could reach approximately 99.75% straight-through processing, with normal transactions completed in around 15-30 minutes and almost no human intervention. These figures reflect practitioner experience, not industry-wide benchmarks.

RPA changed more than cost. It changed operating hours, scalability, accuracy and customer experience. A human works shifts. RPA operates continuously. A global shared-services centre can therefore process work across countries and time zones.

But banking is entering another phase. The question is no longer only: Which task can we automate? It is becoming: Which business outcome can an AI agent manage?

From RPA to GenAI

RPA became successful because banking contains many structured processes. The rules, workflow, screens and sequence are known. If X occurs, the bot performs Y. If validation fails, an exception is raised.

Generative AI introduced another capability. Systems can increasingly interpret unstructured documents, emails, contracts, customer instructions, policies and procedures. This matters because a significant share of banking operations still involves documents and natural language.

Standard Chartered's Trade AI Engine with IBM is an authentic example of this transition. Launched in 2019, it was designed to convert non-digital trade documents into machine-readable information using OCR, machine learning and natural-language processing. Standard Chartered reported processing more than 36 million pages of trade documents annually, containing more than 200 million data elements.

From GenAI to Agentic AI

Generative AI normally responds to a request. Agentic AI can pursue an objective within defined boundaries. An agent can understand a goal, retrieve information, reason about context, select tools, execute permitted actions, observe results and determine the next action.

This moves banking from task automation toward process orchestration and, eventually, bounded autonomous operations. The word bounded is critical. Banking autonomy cannot mean uncontrolled autonomy.

Reimagining LC Issuance

Imagine that a corporate customer submits an LC application, commercial contract, invoice, shipping instructions and supporting documents. The objective becomes: Issue a compliant LC for this customer.

An intake agent can classify documents, extract information, identify missing data and detect inconsistencies. A trade agent can retrieve relevant policies, evaluate LC terms and identify unusual clauses. A compliance agent can assemble evidence about counterparties, countries, transaction patterns and risk indicators.

A credit agent can retrieve approved facilities, available limits, utilization, collateral conditions and expiry dates. A document agent can compare amounts, dates, goods descriptions, counterparty names, shipping information and contractual requirements. An LC drafting agent can prepare the LC structure, propose clauses and flag non-standard language.

Standard Chartered wrote in 2025 that AI can increasingly extract information from bills of lading and invoices and compare extracted information against specific LC terms. That provides useful industry validation for this direction.

RPA Does Not Disappear

A major mistake would be to assume that AI agents simply replace RPA. A stronger architecture is: AI reasons. Rules constrain. APIs transact. RPA bridges legacy. Humans govern exceptions.

An agent may determine what needs to happen. An API, RPA bot, workflow engine, core banking system or human may execute the action. RPA becomes one execution capability available to an orchestration layer.

Existing RPA estates also contain valuable operational knowledge: process flows, business rules, system dependencies, exception patterns, controls and authorization points. Banks should preserve and reuse this knowledge.

Deterministic Banking Must Remain

Banking cannot convert every rule into an AI decision. Transaction limits, authorization mandates, accounting controls, SWIFT validation, regulatory restrictions, system entitlements, segregation of duties and approval thresholds may need deterministic enforcement.

The target architecture is therefore probabilistic intelligence inside deterministic controls. Intelligence can come from AI. Authority must come from the bank.

A Verifiable Agentic Banking Case

JPMorgan researchers developed Legal Agentic Workflows, or LAW, for custody and fund-services contracts. The architecture orchestrated specialized text agents and domain-specific tools rather than relying on a single general-purpose LLM.

For a complex termination-date calculation task, the paper reported a 92.9 percentage-point improvement over its baseline. The transferable lesson is architectural: complex banking work benefits from specialized agents, tools, retrieval, validation and orchestration.

Agent Identity and Governance

Every production agent should have a controlled identity. The bank should be able to reconstruct which agent acted, which model it used, what data it accessed, which tools it called, what permissions it had, what evidence supported its decision and what happened next.

NIST's AI Risk Management Framework provides a useful structure: Govern, Map, Measure and Manage. Agentic banking should operationalize these principles through permissions, auditability, monitoring, escalation and control.

Regulatory expectations are also evolving. In April 2026, U.S. banking regulators issued revised model-risk-management guidance. Agentic architectures require banks to consider model risk together with tool use, data access, permissions and autonomous action.

The Shared-Service Centre Becomes Agentic

RPA transformed shared services once. Agentic AI may transform them again. The future centre may combine human specialists, AI agents, RPA bots, APIs, rules engines and autonomous workflows.

Human roles can shift toward exception management, AI supervision, control management, process architecture, knowledge management, complex customer problems and continuous improvement.

BNY has publicly described 'digital employees' and multi-agent capabilities in its enterprise AI environment, providing evidence that this operating-model direction is moving beyond theory.

A Practical Migration Model

Stage 1 — Map the RPA estate. Identify high-volume, high-cost, exception-heavy, document-heavy, and judgement-heavy processes.

Stage 2 — Decompose each process. Separate deterministic execution from reasoning, knowledge retrieval, and judgement.

Stage 3 — Introduce agents gradually. Start with read-only retrieval, then recommendations, preparation, controlled execution, and finally bounded autonomy.

Stage 4 — Establish controls. Create agent identities, permissions, transaction limits, evidence requirements, logging and escalation.

Stage 5 — Integrate the execution layer. Prefer APIs where practical and retain RPA where legacy systems require it.

Stage 6 — Scale based on evidence. Measure accuracy, turnaround, straight-through processing, cost, exception rates, overrides, policy violations, and customer impact.

The Autonomy Ladder

Level 0 — Manual: humans perform the process.

Level 1 — RPA: bots execute predefined workflows.

Level 2 — AI Assisted: AI interprets information while humans decide.

Level 3 — Agent Assisted: agents reason and recommend while humans authorize key actions.

Level 4 — Bounded Agentic: agents execute independently within defined limits while humans supervise exceptions.

Level 5 — Autonomous Operations: multiple agents manage complete outcomes inside regulatory and control boundaries.

Banks should not jump directly to Level 5. Autonomy should be earned through evidence.

The Next Banking Revolution

RPA demonstrated that enormous amounts of banking work could be automated. We automated actions, workflows and processing. We created digital workers.

The next phase may automate parts of reasoning, coordination and decision preparation. Within controlled limits, banks may eventually automate complete business outcomes.

The winning architecture will probably not be AI replacing RPA. It will be AI + RPA + APIs + Rules + Humans. RPA provides deterministic execution. APIs provide connectivity. Rules provide hard boundaries. AI provides contextual intelligence. Agents provide orchestration. Humans provide judgement and accountability.

RPA automated our hands. AI augments our judgement. Agents orchestrate both.

That is the real journey: from automated processes to intelligent operations.

References

Standard Chartered. “We've pioneered Trade AI Engine with IBM.” 22 May 2019. https://www.sc.com/en/press-release/weve-pioneered-trade-ai-engine-with-ibm/

Standard Chartered. “Augmented Intelligence: A Pragmatic Look at AI in Trade.” 2025. https://www.sc.com/en/news/corporate-investment-banking/augmented-intelligence-a-pragmatic-look-at-ai-in-trade/

Watson et al. “LAW: Legal Agentic Workflows for Custody and Fund Services Contracts.” 2024. https://arxiv.org/abs/2412.11063

JPMorgan Chase. “Securing the Next Generation of AI Agents.” 2026. https://www.jpmorganchase.com/about/technology/blog/securing-agentic-ai

BNY. Annual Report 2025. https://www.bny.com/corporate/global/en/investor-relations/annual-report-2025.html

NIST. AI Risk Management Framework Playbook. https://www.nist.gov/itl/ai-risk-management-framework/nist-ai-rmf-playbook

Federal Reserve. SR 26-2, Revised Guidance on Model Risk Management. April 2026. https://www.federalreserve.gov/supervisionreg/srletters/SR2602.htm

McKinsey & Company. Global Banking Annual Review 2025. https://www.mckinsey.com/industries/financial-services/our-insights/global-banking-annual-review-2025

Read my Article "Has China Changed the Race of AI' ttps://www.linkedin.com/pulse/has-china-changed-ai-race-mohammad-arif-9zmbc/