Building AI Agents for FinTech and Payment Systems: ISO 8583 & NDC+

How AI agents can analyse ISO 8583 transactions, monitor NDC+ ATM networks and support reconciliation and fraud investigation with the right guardrails.

By Mudasser Tufail · · 9 min read

  • FinTech
  • ISO 8583
  • NDC+
  • AI Agents
  • Payments
  • Banking

Introduction

The financial technology industry is entering a new era where Artificial Intelligence is moving beyond chatbots and traditional automation toward intelligent, autonomous AI agents.

Banks, payment processors, financial institutions, and ATM service providers manage millions of transactions every day. Behind these transactions are complex systems responsible for authorization, routing, settlement, reconciliation, fraud detection, and operational monitoring.

Technologies such as ISO 8583 and NCR Direct Connect Plus (NDC+) play an important role in this ecosystem. However, understanding transaction messages, identifying operational issues, investigating failed transactions, and maintaining payment infrastructure still require significant engineering and operational effort.

This is where AI agents can make a meaningful difference.

By combining AI agents with existing banking infrastructure, financial institutions can introduce intelligent automation while maintaining the reliability, security, and accuracy required by financial systems.

Understanding the FinTech Payment Ecosystem

Before exploring AI agents, it is important to understand the technologies that power traditional payment systems.

1. ISO 8583 – The Language of Financial Transactions

ISO 8583 is an international standard that defines the message structure and format used to exchange card-originated financial transaction messages between acquiring institutions and card issuers.

It is commonly used across:

  • Point-of-Sale (POS) terminals.
  • ATM transaction processing.
  • Payment switches.
  • Card authorization systems.
  • Banking and financial transaction networks.

An ISO 8583 message contains different components, including:

  • MTI (Message Type Indicator): Identifies the type of transaction message.
  • Primary and Secondary Bitmap: Indicates which data elements are present.
  • Data Elements: Contains transaction information such as processing code, transaction amount, transmission date, system trace audit number, terminal information, and response code.

For example, a typical card transaction may involve:

  1. A customer initiates a payment through a POS terminal.
  2. The terminal generates an ISO 8583 authorization request.
  3. The payment switch receives and processes the request.
  4. The transaction is routed to the appropriate issuer.
  5. The issuer evaluates the transaction and returns an authorization response.
  6. The response travels back through the switch to the POS terminal.

Although this process appears straightforward, troubleshooting transaction failures can be complicated because multiple systems, message fields, and network components are involved.

2. NDC+ – ATM Communication and Transaction Management

NCR Direct Connect Plus (NDC+) is a proprietary ATM communication protocol used in NCR ATM environments.

It enables communication between ATM terminals and host systems, supporting transaction flows, terminal commands, device status messages, and ATM operational management.

NDC+ is particularly relevant for:

  • ATM transaction processing.
  • Terminal-to-host communication.
  • Host-to-terminal commands.
  • ATM device status monitoring.
  • Electronic journal (EJ) processing.
  • Cash dispensing and transaction workflows.
  • ATM configuration and operational management.

Consider a typical ATM cash withdrawal:

  1. A customer inserts a card.
  2. The ATM collects the required transaction information.
  3. The ATM communicates with the host system.
  4. The host processes the financial authorization.
  5. The ATM receives the appropriate response.
  6. The ATM executes the cash dispensing operation.
  7. The transaction is recorded in the ATM's electronic journal.

An important distinction is that ISO 8583 and NDC+ serve different purposes. ISO 8583 defines financial transaction interchange messages, while NDC+ handles NCR ATM terminal communication and control. They can operate together within the same banking transaction ecosystem.

Why Traditional FinTech Systems Need AI Agents

Traditional banking infrastructure is designed around deterministic processing, predefined business rules, and structured transaction messages.

While these systems are reliable, operational teams still face several challenges.

1. Complex Transaction Investigation

When a transaction fails, engineers often need to inspect multiple ISO 8583 fields, response codes, transaction logs, and switch records to identify the root cause.

2. ATM Monitoring and Operational Issues

ATM networks generate continuous status messages, device errors, communication failures, and operational alerts. Identifying the actual cause of an issue can require manual investigation.

3. Payment Reconciliation

Financial institutions need to compare transaction records across multiple systems to identify missing, duplicated, reversed, or inconsistent transactions.

4. Fraud Detection

Traditional rule-based systems can identify predefined suspicious patterns, but emerging fraud techniques require more adaptive investigation and contextual analysis.

5. Legacy System Complexity

Many financial institutions operate established banking platforms that have evolved over several years. Integrating new capabilities without disrupting existing infrastructure requires careful planning.

AI agents can help address these challenges by connecting to existing systems, gathering relevant information, analyzing patterns, and executing approved workflows.

How AI Agents Can Transform FinTech and Payment Systems

Unlike traditional chatbots that primarily respond to questions, AI agents can use tools, access authorized APIs, coordinate multiple steps, and work toward defined objectives.

Here are some practical applications.

1. Intelligent ISO 8583 Transaction Analysis Agent

An ISO 8583 analysis agent can inspect transaction messages and explain their structure and processing outcomes.

For example, an engineer could provide a transaction message and ask:

"Why was this transaction declined?"

The agent could:

  • Parse the MTI and bitmap.
  • Extract relevant data elements.
  • Identify the processing code.
  • Analyze the response code.
  • Compare the request and response messages.
  • Retrieve related transaction records.
  • Identify potential inconsistencies.
  • Generate a human-readable investigation report.

Instead of manually examining multiple fields, engineers can interact with the agent using natural language.

2. ATM Monitoring and NDC+ Analysis Agent

An NDC+ monitoring agent can continuously analyze ATM messages, terminal status information, and electronic journals.

Potential capabilities include:

  • Identifying ATM communication failures.
  • Detecting repeated device errors.
  • Monitoring cash dispenser issues.
  • Analyzing ATM status messages.
  • Identifying unusual transaction patterns.
  • Correlating terminal errors with host-side transaction failures.
  • Generating operational incident reports.

For example, if an ATM repeatedly reports cash dispenser errors, the agent can correlate the messages with previous incidents and relevant operational records.

It can then provide a probable root cause and recommend the next troubleshooting steps.

3. Intelligent Payment Reconciliation Agent

Reconciliation is one of the most valuable areas for AI-driven automation.

An AI agent can compare transaction records from different systems, identify discrepancies, and prepare investigation reports.

For example:

  • Compare switch transactions with core banking records.
  • Identify transactions missing from settlement files.
  • Detect duplicate transaction references.
  • Identify mismatched transaction amounts.
  • Investigate reversal-related discrepancies.
  • Categorize reconciliation exceptions.

The agent can gather supporting evidence and prepare a report for the finance or operations team.

Actual financial adjustments should remain subject to deterministic validation and authorized approval.

4. AI-Powered Fraud Investigation Agent

AI agents can support fraud investigation by analyzing transaction history, customer activity, terminal information, and suspicious transaction patterns.

An investigation workflow could include:

  1. Receive a suspicious transaction alert.
  2. Retrieve related transaction history.
  3. Analyze transaction frequency and amounts.
  4. Identify unusual terminal or geographical patterns.
  5. Compare activity with established risk indicators.
  6. Generate an investigation summary.
  7. Escalate the case for human review.

Rather than simply generating another fraud alert, the agent can help investigators understand why a transaction appears suspicious.

5. Payment Operations and Incident Management Agent

Financial institutions operate complex payment infrastructures involving switches, APIs, databases, networks, and external service providers.

An operations agent can integrate with monitoring platforms and incident management systems to:

  • Analyze payment processing failures.
  • Investigate API errors.
  • Identify recurring infrastructure problems.
  • Correlate logs across services.
  • Summarize system health.
  • Generate incident reports.
  • Recommend remediation steps.

With appropriate permissions, it can also initiate approved operational workflows.

Designing a Multi-Agent Architecture for FinTech

A single AI agent does not necessarily need to handle every banking operation.

A more scalable approach is to build a multi-agent architecture in which specialized agents are responsible for specific business domains.

A possible architecture could include:

1. Transaction Analysis Agent

Responsible for parsing ISO 8583 messages, analyzing transaction fields, and investigating authorization failures.

2. ATM Operations Agent

Responsible for analyzing NDC+ messages, ATM device status, electronic journals, and terminal communication.

3. Fraud Investigation Agent

Responsible for analyzing suspicious transaction patterns and gathering supporting evidence.

4. Reconciliation Agent

Responsible for comparing transaction records and identifying financial discrepancies.

5. Monitoring Agent

Responsible for monitoring system health, processing alerts, and identifying operational anomalies.

6. Orchestrator Agent

Coordinates specialized agents, manages workflows, combines findings, and produces the final investigation result.

These agents can communicate through a controlled orchestration layer and access banking systems through authorized tools and APIs.

Reference Architecture

A simplified architecture can be organized into five layers:

Layer 1: Existing Banking Infrastructure

  • ATM networks.
  • POS terminals.
  • Payment switches.
  • Core banking systems.
  • ISO 8583 interfaces.
  • NDC+ communication infrastructure.
  • Transaction databases.

Layer 2: Integration and Data Access

  • API integrations.
  • Message consumers.
  • Event streaming.
  • Log aggregation.
  • Secure database access.
  • Transaction data normalization.

Layer 3: AI Agent Platform

  • Transaction analysis agent.
  • ATM monitoring agent.
  • Fraud investigation agent.
  • Reconciliation agent.
  • Operations agent.
  • Agent orchestration.

Layer 4: Knowledge and Context

  • ISO 8583 message specifications.
  • NDC+ protocol documentation.
  • Banking business rules.
  • Historical transaction records.
  • Troubleshooting procedures.
  • Operational knowledge base.

A Retrieval-Augmented Generation (RAG) architecture can help agents retrieve relevant technical documentation and operational knowledge instead of relying entirely on the model's internal knowledge.

Layer 5: User Interface and Reporting

  • Natural language interface.
  • Transaction investigation dashboard.
  • ATM monitoring dashboard.
  • Incident management.
  • Reconciliation reports.
  • Audit and approval workflows.

Important Considerations When Building AI Agents for Banking

Financial systems have requirements that are significantly different from ordinary business applications.

AI agents must be designed with these considerations from the beginning.

1. Security and Data Protection

Sensitive financial information must be protected. Cardholder data, PIN-related information, authentication credentials, and other restricted information must never be unnecessarily exposed to AI models.

Access should follow least-privilege principles, with appropriate masking, encryption, and data retention controls.

2. Deterministic Transaction Processing

AI models are probabilistic, whereas payment processing requires predictable and verifiable outcomes.

An AI agent should not independently determine whether a financial transaction is authorized or directly modify transaction records based on an LLM-generated decision.

Financial authorization, message validation, settlement, and other critical operations must remain within deterministic, properly controlled systems.

3. Human Approval and Governance

High-impact operations should require appropriate authorization.

For example, an agent may identify a suspicious transaction and recommend blocking a card, but the actual action should follow the bank's approved authorization policy.

4. Auditability and Traceability

Every agent action should be traceable, including:

  • Input transaction references.
  • Retrieved information.
  • Tools and APIs accessed.
  • Agent-generated recommendations.
  • Approval decisions.
  • Actions executed.
  • Final outcomes.

5. Reliability and Continuous Evaluation

AI agents should be evaluated against realistic transaction scenarios, historical incidents, edge cases, and failure conditions.

Testing should include incorrect messages, network failures, duplicate requests, timeouts, reversals, and partial system failures.

My Experience in FinTech and Banking Technology

My experience in software engineering includes working on financial technology solutions across multiple domains, including POS systems, mobile banking, ATM monitoring, and enterprise financial software.

I have also had the opportunity to contribute to projects involving a renowned banking institution, gaining exposure to the engineering challenges associated with financial systems and banking operations.

Working with these systems provides an understanding of the importance of transaction accuracy, system reliability, integration complexity, security, and operational visibility.

This experience is particularly relevant when exploring how AI agents can be introduced into existing banking infrastructure.

The objective should not be to replace established financial systems with AI. Instead, it should be to enhance existing infrastructure with intelligent capabilities that help engineering, operations, and business teams work more efficiently.

The Future of Agentic AI in FinTech

The future of financial technology will likely involve a combination of traditional transaction-processing infrastructure and intelligent AI-driven automation.

ISO 8583, ATM communication protocols, payment switches, and core banking platforms will continue to provide the deterministic foundation required for financial transactions.

AI agents can build on top of this foundation by introducing intelligent investigation, operational automation, contextual analysis, and workflow orchestration.

The opportunity is not simply to add an AI chatbot to a banking application. It is to build secure, specialized, and well-governed AI agents that understand financial workflows and interact with existing systems through controlled interfaces.

For FinTech companies and banking institutions, this represents an opportunity to modernize operations without abandoning the reliability of established payment infrastructure.

The next generation of FinTech systems will not just process transactions. They will also intelligently understand, investigate, monitor, and optimize the operations surrounding those transactions.


Related: my work as a FinTech software consultant, the POS and ATM monitoring (EPM) case study, and building AI agents as an agentic AI engineer.

Let's work together

Looking to discuss a project or explore potential collaboration? Reach out anytime.