Introduction
AI agents in banking are moving beyond demonstrations and pilot projects as major financial institutions deploy them in payment repair, trade processing, document checking, liquidity management and client onboarding. At Sibos 2026 in Miami, banks including BNY, BNP Paribas, Deutsche Bank, Citi and HSBC described production or expanding deployments in which artificial intelligence performs defined operational tasks while human employees retain responsibility for material decisions.
The developments highlight a broader change in how financial institutions are approaching artificial intelligence. Rather than treating AI as a standalone technology project, banks are increasingly connecting agents to existing workflows, data systems and control frameworks. The approach also reflects the constraints of regulated financial services, where errors involving payments, customer information or financial decisions can have operational and compliance consequences.
At the same time, adoption remains uneven. Bottomline’s sixth Payments Intelligence Gap study found that more than a third of the 300-plus banking and payments professionals surveyed had no AI in their payment operations. Meanwhile, almost half identified AI integration as a critical or high priority for the following 12 months.
What AI Agents Mean for Banking Operations
AI agents differ from conventional software automation because they can interpret information, determine the next step within an assigned workflow and interact with multiple systems rather than simply executing a fixed instruction.
For banks, this distinction is important because many operational processes contain unstructured information and exceptions. Payment instructions can contain incorrect beneficiary information, while trade transactions may involve large volumes of documents and conditions that require interpretation.
At BNY, Chief Data and AI Officer Sarthak Pattanaik described a digital employee designed to handle payment-repair cases. The system reads payment fields, checks syntax against ISO 20022 and evaluates whether the transaction is consistent with the stated movement of value. Deterministic controls provide safeguards, while employees retain responsibility for judgement calls. BNY has been developing its Eliza platform since 2023.
This model illustrates a recurring structure across the banking deployments discussed at Sibos: AI performs defined reasoning or processing tasks, while rules and human oversight establish the boundaries within which the technology can operate.
How Banks Are Deploying AI Agents
Payment repair and exception handling
Payments remain one of the clearest areas for agentic automation because banks process large volumes of transactions while a smaller proportion require manual intervention.
BNY is using AI to examine payment exceptions involving fields such as beneficiaries, addresses, references and payment purposes. The technology is combined with ISO 20022 validation and deterministic controls before human judgement is applied to cases requiring escalation.
Finastra also announced Repair Recommendations for its OperatorAssist product. The system is designed to identify payment exceptions, determine potential causes and recommend repairs while accounting for payment frameworks including Swift, Fedwire, SEPA, UPI and Nexus.
Securities and trade processing
BNP Paribas is applying agentic AI to securities-services trade processing. Chief AI Officer for Corporate and Institutional Banking Charles Holive said a process had been reduced from ten steps to six, with an AI agent dispatching work between stages.
The bank also reported an example involving email triage. When the system initially classified almost every incoming message as an exception, employees used those exceptions as opportunities to train the agent. Within a couple of weeks, the system was handling 80 to 85 per cent of the workload.
BNP Paribas said its AI programme is organized around six major transformation programmes under its plan to 2030 rather than a large number of isolated proofs of concept. Model interactions are processed through an observability layer with controls established by compliance, legal and human-resources functions.
Lending and financial research
Deutsche Bank has also moved toward reusable agentic infrastructure through a framework called Ada. Joanne Hannaford, CIO and CPO of the bank’s Corporate Bank, said agents are being deployed at industrial scale.
The bank was a design partner for Google Cloud’s financial research agent within Gemini Enterprise for Financial Services. The system provides coverage bankers with information relating to markets, macroeconomic conditions and clients.
Data infrastructure is another component of the approach. Deutsche Bank’s German lending business used knowledge graphs to support a new small-business client onboarding and lending process that was completed in a day, compared with a process that previously would have taken a month.
Client onboarding and treasury
Citi is developing agent-based approaches to onboarding documents and allowing transactions to move toward agent-to-agent exchanges between banks and corporate clients. The bank is rolling out the approach in North America.
Citi executive Stephen Randall also pointed to Arc, the enterprise-wide control layer discussed by Chief Executive Jane Fraser during the Sibos plenary. The bank is working with regulators on AI use cases, including whether people remain in the loop, supervise the process or are removed from individual stages of a workflow.
AI agents are also being used within Google’s own finance operation. Vice President of Finance Kristin Reinke said its treasury function processes well over 100,000 transactions a day and uses agents to forecast surplus cash and identify where it could be placed, while the cash manager retains the final decision.
Human Oversight Remains Central to Banking AI
The banking deployments presented at Sibos share an important characteristic: automation does not necessarily mean removing people from the process.
Instead, banks are assigning AI systems to repetitive or information-intensive stages while reserving decisions with financial, regulatory or customer consequences for employees.
This distinction is particularly relevant in financial services because AI systems can produce incorrect outputs even when they perform well across large datasets. BNP Paribas’ experience with email triage illustrates the issue. The initial system sent almost everything to exception handling, requiring employees to turn those failures into training inputs.
BNP Paribas also highlighted the importance of the technology surrounding an AI model. Holive said the bank’s focus was on improving the “harness” around models, including governance and controls, rather than waiting for progressively newer models.
This suggests that operational reliability depends not only on model capability but also on access controls, monitoring, data quality, workflow design and escalation procedures.
Customer Acceptance Creates Another Constraint
Bank customers appear more comfortable with conversational AI than with systems that independently execute financial actions.
Research released by Temenos and Celent at Sibos surveyed more than 2,500 banking customers across five regions. It found that 68 per cent would use a conversational AI interface for banking queries, while fewer than half would definitely allow AI to take actions for them. Privacy and data security were cited as concerns by 47 per cent, while 36 per cent identified the risk of AI errors.
The findings provide an important distinction for financial institutions. Customers may accept AI as an interface for obtaining information without necessarily granting the technology authority to initiate transactions or make decisions on their behalf.
The same research found that three in four customers were only moderately satisfied or less with their main bank, more than half were dissatisfied with payment services and a quarter had recently considered switching. Separately, 46 per cent of the banks surveyed by Celent planned major core-system changes in 2027.
These figures place AI adoption within a wider technology and customer-service transition rather than treating it as an isolated innovation.
The Expansion of AI Into Trade Finance
Trade finance is another major area for agentic AI because the sector continues to depend heavily on documents.
HSBC Chief Operating Officer of Global Trade Solutions Tom Elliott said around four billion trade documents move every day, with only one to two per cent being digital. The bank’s Smart Checking system divides the process into modules covering document digitisation and reasoning about transaction conditions.
The system incorporates the rules of the International Chamber of Commerce and the specialist knowledge of HSBC’s trade professionals. Human specialists remain involved in exceptions, higher-value transactions and situations where the system lacks sufficient confidence.
The example demonstrates why trade finance is potentially suitable for AI-assisted workflows while also showing the limitations of full automation. Documentation can be structured and analysed by software, but unusual transactions and regulatory or commercial exceptions may still require specialist interpretation.
AI Adoption Remains Uneven Across Financial Institutions
The emergence of production deployments at large banks does not mean the broader banking industry has reached the same level of adoption.
Bottomline’s survey of more than 300 banking and payments professionals found that more than a third had no AI in their payment operations. At the same time, nearly half regarded integration as a critical or high priority for the next 12 months.
The difference between these positions reflects several practical barriers. Banks must integrate AI with legacy infrastructure, establish data controls, determine responsibility for AI-generated decisions and demonstrate appropriate oversight to regulators.
The result is likely to be a varied adoption landscape, with larger institutions deploying agents in narrowly defined production environments while other organizations remain in evaluation or early implementation stages.
Cross-Border Payments Are Developing Alongside AI
AI was only one part of the technology developments reported on the second day of Sibos 2026.
Citi became the first bank to launch multi-market instant payments through the Swift payments scheme. The service allows clients to access domestic instant-payment rails in Australia, the UK, India and the US through one account structure without separate local banking relationships.
Swift said more than 100 institutions were live or going live on its consumer payments framework. It also reported that most of its 17 first-mover banks had used its ledger for round-the-clock payments in five currencies: euro, sterling, Hong Kong dollar, Singapore dollar and US dollar.
The developments show that banks are working simultaneously on the intelligence applied to financial workflows and the infrastructure used to move money across borders.
Tokenisation Adds Another Layer of Financial Infrastructure
Digital assets and tokenised financial instruments were also part of the Sibos discussions.
Wayne Hughes, who leads the central digital assets team in BNP Paribas’ Securities Services business, said client interest in digital assets was increasing but that only a small number of clients were ready for production deployment.
Tokenised money market funds were identified as an early use case, particularly for distribution, while collateral was described as a longer-term area of interest. Hughes said central bank money, tokenised deposits and stablecoins could each have a role in future financial infrastructure.
He also identified cross-border regulation as a major unresolved issue. A transaction involving a client in Asia-Pacific purchasing a tokenised asset in Europe can raise questions about which rules apply across jurisdictions.
That issue is separate from the technical ability to transfer digital assets. Interoperability, legal recognition, custody arrangements and regulatory coordination remain important components of institutional adoption.
Risks and Limitations of AI Agents in Banking
The expansion of AI agents introduces several operational considerations.
Model reliability: An AI system can produce incorrect or inappropriate outputs, making validation and escalation mechanisms important in regulated workflows.
Data quality: Agent performance depends on the accuracy, structure and accessibility of the data connected to the workflow. Deutsche Bank’s use of knowledge graphs illustrates the importance of data infrastructure alongside AI models.
Governance: Banks need to establish where an AI system can act independently and where human authorization remains necessary.
Legacy integration: Connecting AI to established banking systems can require significant technology changes, particularly when institutions operate complex core infrastructure.
Customer trust: Temenos and Celent’s research indicates that customers distinguish between using AI to obtain information and allowing AI to take actions on their behalf.
Regulatory oversight: Banks are increasingly discussing AI use cases with supervisors, including the role of humans within automated workflows.
These constraints mean that deploying an AI model is only one part of implementing an agentic banking system.
Future Outlook for AI Agents in Banking
The developments presented at Sibos 2026 indicate that the banking industry’s focus is shifting from whether AI can perform individual tasks toward how those tasks can be integrated safely into production operations.
Payment repair, trade-document analysis, securities processing, customer onboarding, financial research and treasury management are among the areas where banks are testing or deploying agent-based workflows.
The next stage will depend on factors beyond model performance. Governance frameworks, data architecture, system integration, regulatory engagement and customer acceptance are likely to influence how extensively AI agents can operate within financial institutions.
The simultaneous development of instant-payment infrastructure, tokenised assets and AI-driven workflows also points to a broader restructuring of financial operations. These technologies address different parts of the banking system, but their effectiveness depends on interoperability, common standards and controls across institutions and jurisdictions.
For now, the evidence from Sibos 2026 points to a hybrid operating model: AI increasingly handles defined operational work, while people retain responsibility for exceptions, oversight and decisions carrying greater financial or regulatory consequences.
Conclusion
AI agents in banking are moving into production across selected financial workflows, but adoption remains uneven across the industry. BNY, BNP Paribas, Deutsche Bank, Citi and HSBC have described deployments covering payments, securities, trade finance, onboarding and other operational functions.
The emerging model is not simply the replacement of employees with autonomous software. Instead, banks are combining AI reasoning with deterministic controls, monitoring systems and human judgement. Customer research and industry surveys also show that acceptance and implementation vary significantly.
As banks continue to modernize payment infrastructure and explore tokenisation alongside AI, the principal issue is increasingly how these technologies can be integrated into regulated financial systems with appropriate controls, rather than whether the underlying technology can perform individual tasks.
