AI Agents for Banking: How Banks Are Moving Beyond Chatbots

For years, artificial intelligence in banking largely worked behind the scenes.

Algorithms detected suspicious transactions, assessed risk, recommended products, screened documents and helped banks analyse enormous amounts of customer data.

Then generative AI changed the interface.

Suddenly, customers and employees could talk to AI systems in natural language, ask complicated questions and receive surprisingly sophisticated answers.

Now another shift is beginning.

Instead of merely answering a question, an AI system may be able to determine what needs to happen next, access approved tools, complete several steps and execute an action within predefined boundaries.

That is the promise of AI agents.

Consider the difference.

A traditional banking chatbot might answer:

“Here is how you replace your credit card.”

An AI agent could potentially understand that the card has been lost, verify the authenticated customer’s instruction, block the existing card, initiate a replacement, update the customer on delivery and escalate unusual circumstances to a human employee.

The difference is subtle but fundamental.

One system provides information.

The other participates in the workflow.

That transition has already begun. In July 2026, DBS announced that its corporate banking assistant DBS Joy had become fully agentic in Singapore. Authenticated corporate and SME customers can move beyond asking questions and complete selected everyday banking tasks in the same conversation. DBS said its AI-enabled assistants collectively reach more than 10 million customers across Singapore, Hong Kong and Taiwan.

However, banking is not an ordinary software environment.

When AI begins interacting with payments, customer records, compliance processes, credit decisions or investments, a wrong answer can become a wrong action.

Therefore, the most important question is not simply whether banks can deploy AI agents.

It is whether they can give software greater autonomy without giving up accountability.

AI Agents for Banking: From Answers to Actions

An AI agent is different from a conventional chatbot because it is designed to work toward a goal rather than only generate a response.

The Bank of England describes agentic AI systems as systems capable of taking autonomous actions toward specified goals by using tools, learning from feedback and adapting to changing environments.

In practical banking applications, an agent may combine several capabilities:

understanding natural-language instructions;
retrieving information from approved databases;
planning a sequence of tasks;
calling internal software tools or APIs;
checking whether required information is missing;
coordinating with other specialised agents;
executing permitted actions;
recording what happened;
escalating exceptions to humans.

Consequently, the shift from generative AI to agentic AI can be thought of as the difference between knowing and doing.

Generative AI might draft a KYC summary.

An agentic system could gather the documents, extract relevant information, identify gaps, request follow-up information and prepare the case for human approval.

That difference creates significant economic potential.

It also increases risk.

Why Banking Is Particularly Suitable for Agentic Workflows

Banks contain thousands of processes built around information.

Employees routinely gather documents, compare records, verify information, move data between systems, investigate anomalies, prepare reports, answer customer questions and escalate cases according to specific rules.

Many of these processes are not simple enough for traditional automation.

They contain exceptions, require context.

They may involve unstructured documents or conversations.

That is exactly where more capable AI systems become interesting.

McKinsey’s 2025 research into Asian banking operations identified multiple areas where reusable AI agents could potentially work across banking workflows, arguing that agentic systems could reshape areas ranging from customer operations to servicing and risk-related processes. However, the consultancy also notes that extracting value requires banks to redesign workflows rather than simply placing an AI layer over existing processes.

This distinction matters.

Automating a bad process simply creates a faster bad process.

The larger opportunity is to reconsider how the work should happen when software can reason across multiple steps.

Customer Service Is Becoming Transactional, Not Just Conversational

Customer service is one of the easiest places to see the difference between chatbots and agents.

Traditional banking bots are generally built around questions such as:

“What is my card limit?”

“How do international transfers work?”

“Where can I download a statement?”

AI agents can potentially move the interaction one step further:

“Show me my card spending this month and replace the card I lost yesterday.”

Now the system has to understand multiple intentions, retrieve information, follow security requirements and execute actions.

DBS is already moving in this direction.

Its July 2026 announcement said DBS Joy had become agentic for its approximately 350,000 corporate users, while its retail DBS digibot was being prepared for selected agentic capabilities such as checking card usage, calculating reward points, requesting fee waivers and blocking or replacing cards. The bank said these agentic functions would operate only after customers authenticate themselves and would act on instructions initiated by those customers.

Importantly, humans have not disappeared from the system.

DBS maintains escalation routes to human customer-service staff for more complicated needs.

That hybrid model may become one of the dominant patterns in banking:

AI handles routine execution → humans handle ambiguity, judgement and sensitive exceptions.

KYC and Compliance May Be One of the Strongest Early Use Cases

Know-Your-Customer processes illustrate why agentic AI can be more valuable than another customer-facing chatbot.

KYC work is essential, yet it often involves significant manual effort.

Employees may need to collect customer records, review supporting evidence, compare sources, identify missing information and prepare documentation for approval.

Bank of Singapore, OCBC’s private banking arm, has already applied an agentic system to its Source of Wealth process.

According to OCBC’s 2025 annual report, the system independently coordinates activities including extracting and structuring information from customer records and documents, synthesising evidence, identifying information gaps and prompting targeted follow-ups.

The reported efficiency improvement is substantial.

OCBC says work that previously could take up to 10 days can now be completed in about one hour. Yet relationship managers and compliance teams retain responsibility for judgement and final approval.

That final detail is more important than the headline efficiency number.

The system automates process complexity without outsourcing accountability.

Why Compliance Work Fits the Agent Model

Many compliance processes contain four ingredients well suited to AI agents:

large volumes of documents;
repetitive information gathering;
defined procedural requirements;
clear points where human judgement can be inserted.

This makes compliance an interesting testing ground for more autonomous AI.

The Bank for International Settlements has also examined AI agents as potential copilots for compliance work, including assisting with investigations and routine work involved in preparing suspicious-activity reports.

However, financial institutions still need robust validation because incomplete or fabricated information in compliance workflows can create serious consequences.

Fraud Detection Could Evolve Into Fraud Investigation

Banks have used machine learning for fraud detection for years.

Agentic AI changes the potential scope of that work.

A traditional fraud model might generate a risk score:

Transaction X has an 87% probability of being suspicious.

An AI agent could potentially investigate why.

For instance, it might gather previous account behaviour, compare devices, analyse transaction patterns, identify related accounts, retrieve relevant policies and assemble an investigation summary before sending the case to a fraud specialist.

This could reduce the amount of time specialists spend collecting information.

However, fraud is also an example of why autonomy must be bounded.

Automatically freezing a customer’s account based on an unreliable AI conclusion could produce significant financial and reputational harm.

Therefore, the most valuable role of an agent may initially be investigation and orchestration rather than unrestricted decision-making.

The Bank of England’s assessment of AI adoption in financial services similarly identifies combating financial crime alongside internal process optimisation and customer support as major near-term AI use cases.

Relationship Managers Could Gain AI Research Teams

Private bankers and corporate relationship managers spend considerable time preparing for client conversations.

They may need to understand:

account activity;
portfolio performance;
relevant products;
corporate financial information;
market developments;
previous conversations;
upcoming maturities;
customer service issues.

A collection of specialised agents could assemble this information automatically.

Instead of starting with twenty different systems, a relationship manager might begin the day with an organised brief:

“Three clients require attention today. One has significant cash reaching maturity. One has unusual account activity that requires review. One may be affected by a market development.”

This moves AI from information retrieval toward decision support.

OCBC is already experimenting with this broader direction in wealth management. Its 2026 OCBC WoW platform architecture incorporates AI agents alongside real-time market data, OCBC research, customer portfolio information and deterministic guardrails to provide always-on wealth capabilities.

The phrase “deterministic guardrails” is especially relevant.

In financial services, some decisions should not be probabilistic.

Credit Workflows Could Become Faster—But Should They Become Autonomous?

Credit is one of the most economically valuable and sensitive banking activities.

An AI system could potentially gather financial statements, identify relevant ratios, analyse repayment behaviour, compare industry trends and prepare a credit memorandum.

DBS said in its 2025 annual report that it had already improved employee efficiency in credit evaluation by using generative AI to extract and validate information, while progressing toward agentic workflows including credit-evaluation memo generation for staff review.

Notice the final phrase: for staff review.

There is an important difference between using AI to prepare a decision and giving AI authority to make that decision.

Credit decisions may have regulatory, fairness, explainability and financial consequences.

The BIS Innovation Hub’s Project Noor reflects these concerns. The project focuses on tools that could help supervisors examine AI models used by financial institutions for transparency, fairness and robustness, noting that AI already influences activities such as mortgage approvals, card limits and fraud detection.

Therefore, greater agent autonomy does not necessarily mean every banking decision should become autonomous.

Payments May Eventually Be Initiated by Machines

One of the most disruptive implications of agentic AI extends beyond banking operations.

It concerns who initiates commerce.

Imagine telling a personal AI:

“Book my business trip next month, stay within my S$3,000 budget and use whichever card gives me the best combination of points and insurance.”

The agent would potentially compare flights, hotels and payment options before completing the purchase.

At that point, the AI is no longer only recommending what to buy.

It is participating in the transaction.

This is already moving from theory toward experimentation.

In February 2026, DBS announced that it was working with Visa on a pilot involving agent-initiated payments through Visa Intelligent Commerce. The initiative explores AI-ready payment credentials, authentication and payment signals designed to enable consent-driven transactions performed by agents on behalf of customers.

This could eventually reshape competition between banks.

Today, consumers often choose which card to use.

Tomorrow, an AI agent may choose according to interest rates, rewards, fees, merchant offers and customer preferences.

That means banks may eventually need to market financial products not only to people—but also to software acting for people.

The Agentic Bank May Be Built From Multiple Specialised Agents

The long-term model may not be one giant AI running an entire bank.

A more realistic architecture is a collection of specialised agents operating within defined permissions.

For example:

Customer Agent
Understands the customer’s request.

Identity Agent
Confirms authentication and permissions.

Product Agent
Retrieves relevant product rules.

Risk Agent
Checks whether the requested action meets risk thresholds.

Transaction Agent
Executes the approved instruction.

Compliance Agent
Reviews applicable regulatory or KYC requirements.

Escalation Agent
Routes unusual cases to a human.

Together, they form a multi-agent workflow.

This structure may provide more control because every agent can have limited responsibilities, tools and permissions.

McKinsey describes a similar direction in its analysis of banking operations, where reusable specialised agents can collaborate across journeys instead of every department building isolated AI applications.

However, more agents also create more interaction points where failures can occur.

That is why orchestration becomes as important as intelligence.

The Biggest Banking AI Problem Is Not Hallucination Alone

Hallucinations receive most of the attention in discussions about generative AI.

They are important—but banking agents introduce a broader risk surface.

A chatbot hallucinating an answer is problematic.

An agent hallucinating and then acting on that incorrect assumption is potentially much more serious.

Key risks include:

Data and Privacy Risk

Banking systems contain highly sensitive personal and financial information.

Agents should only access data required for their assigned task.

Permission Risk

An agent capable of reading an account does not automatically need permission to transfer money.

Tool access should reflect the principle of least privilege.

Model Risk

AI outputs can vary, contain inaccuracies or behave unpredictably in unusual situations.

Cybersecurity Risk

An agent connected to internal tools can become an attractive attack surface.

A malicious instruction that manipulates an agent could have consequences beyond producing incorrect text.

Third-Party Concentration

Banks increasingly rely on external cloud, model and technology providers. Heavy dependence on a small number of providers can create operational concentration risks.

Accountability Risk

If three AI agents collaborate on a process and produce a bad outcome, responsibility must still be traceable.

The BIS has emphasised risks including data security, confidentiality, hallucinations and reputational consequences when financial authorities adopt AI.

Meanwhile, the Association of Banks in Singapore published an updated Handbook on Generative AI Guardrails in Banking in March 2026, covering lifecycle controls and risk-based guardrails for bank deployments.

The key principle is clear:

More autonomy requires stronger controls—not weaker ones.

Human-in-the-Loop May Become Human-at-the-Control-Point

“Human-in-the-loop” is often used as a generic reassurance whenever AI risk is discussed.

But banks need something more specific.

Humans do not need to manually approve every harmless AI action.

That would eliminate much of the efficiency benefit.

Instead, human intervention can be placed according to risk.

Low-Risk Action

Example: retrieving reward-point information.

AI may be permitted to complete the task directly.

Medium-Risk Action

Example: preparing a KYC assessment.

AI completes the workflow, while a human reviews the conclusion.

High-Risk Action

Example: approving a major corporate credit facility.

AI may gather information and recommend an outcome, while authorised humans retain decision authority.

This is essentially risk-tiered autonomy.

The higher the financial, legal or customer consequence, the stronger the approval requirement should become.

OCBC’s Source of Wealth implementation offers a practical example: the agent coordinates much of the preparation, while relationship managers and compliance professionals retain final judgement.

AI Agents Could Create New Financial-System Risks

There is another dimension beyond the individual bank.

What happens if many financial institutions use similar models?

AI agents trained on similar information could respond to market events in similar ways.

If thousands of agents interpret the same signal and make similar decisions simultaneously, AI could potentially amplify market behaviour.

This is still an emerging question rather than an established systemic problem.

In June 2026, the BIS Innovation Hub launched Project Logos with the Bank of England, Deutsche Bundesbank and other partners specifically to study the behaviour of LLM-based portfolio agents in simulated financial markets, including whether shared AI infrastructure could increase homogeneity in financial decision-making.

The Bank of England’s Financial Policy Committee has so far found little evidence that advanced AI has been adopted for core financial decisions at a scale that currently presents systemic risk. Nevertheless, it continues to monitor more advanced generative and agentic applications.

That nuance is essential.

AI agents may transform finance.

But transformation should not be confused with maturity.

What Banks Should Solve Before Scaling Agentic AI

A bank does not become agentic simply by purchasing an AI platform.

Several foundations need to exist first.

1. Clean and Accessible Data

Agents need reliable information.

If customer data are fragmented across incompatible systems, AI will inherit the fragmentation.

2. Defined Tool Permissions

Every agent needs explicit rules regarding what it may read, modify and execute.

3. Traceability

Banks need to know what information an agent accessed, what reasoning path or workflow it followed, which tools it called and what actions occurred.

4. Evaluation

Accuracy must be tested across normal situations, edge cases and adversarial scenarios.

5. Escalation Design

The AI needs to recognise when it should stop.

Knowing when not to act may become one of the most important capabilities of a banking agent.

6. Business Accountability

A named business owner should remain responsible for each agentic workflow.

The agent cannot own the risk.

7. Measurable Economics

Banks should compare the new system against operational outcomes such as:

handling time;
cost per case;
customer satisfaction;
error rates;
fraud losses;
KYC turnaround time;
employee productivity;
conversion;
complaint rates.

AI activity is not the same thing as AI value.

What Investors and Business Leaders Should Watch

Agentic AI could influence banking economics in several ways.

First, it may reduce the cost of servicing routine customer requests.

Second, it could increase relationship-manager capacity by automating information gathering and administrative work.

Third, faster KYC and onboarding may improve customer conversion.

Fourth, better orchestration could reduce operational bottlenecks.

Finally, agent-initiated commerce may change the competitive dynamics of cards, payments and product distribution.

However, businesses and investors should look beyond announcements about the number of AI agents deployed.

From AI Copilot to Agentic Banking Infrastructure

The banking industry has already experienced several generations of automation.

First came deterministic software.

Then machine-learning models,conversational AI.

Then generative AI copilots.

Agentic AI represents another step because the system can increasingly interact with the environment rather than merely analyse it.

DBS describes this direction as moving from AI that helps customers find information toward AI that helps them “get things done.” Its 2026 roll-out provides one of the clearest real-world examples of that transition.

Yet the future is unlikely to be a completely autonomous bank.

Banking depends too heavily on trust, judgement, accountability and regulation.

A more plausible destination is a deeply automated bank where machines perform more of the repetitive orchestration while people maintain responsibility for consequential decisions.

That model is less dramatic than the idea of replacing bankers with autonomous software.

It may also be far more valuable.

The Real Opportunity Is Controlled Autonomy

AI agents could become one of the most important shifts in banking technology because they change the role of artificial intelligence.

AI is moving from analysing.

To recommending.

And increasingly, to acting.

Early deployments already demonstrate the potential.

DBS is allowing agentic systems to complete selected banking workflows after customer authentication. OCBC’s private-banking arm has used an agentic KYC solution to compress parts of a Source of Wealth workflow from as much as 10 days to around an hour. DBS and Visa are even testing infrastructure for transactions initiated by AI agents.

Nevertheless, the central challenge is not building an AI capable of acting.

It is deciding where it should be allowed to act, under what conditions, with what information and who remains accountable when something goes wrong.

That is why the future of AI Agents for Banking will likely be defined by controlled autonomy. Routine tasks may become increasingly machine-operated. Complex cases may become AI-prepared and human-decided.

High-consequence decisions will require stronger approval and governance. The banks that create the most value will therefore not necessarily be those that hand the most authority to AI.

They will be those that know precisely how much authority to give it. And that may become one of the defining capabilities of the next generation of banking.

Technology

Banking’s Next AI Shift: From Answering Questions to Getting Things Done