Gen AI in Singapore’s Banking Sector: Lessons from DBS, UOB, and OCBC

Executive Summary

While Generative AI (Gen AI) has been a hot topic for the past two years (the release of ChatGPT feels like a decade ago), even traditional industries have been proactive in experimenting with and implementing this technology.

In this deep dive, we explore the regulatory environment surrounding Gen AI in the banking industry and analyze how three major banks in Singapore- DBS, UOB, and OCBC- are leveraging this new technology. We also provide recommendations on how enterprises can use Gen AI responsibly and suggest frameworks for integrating it into their workflows.

Industry Context and Market Trends

Monetary Authority of Singapore (MAS) has been active in exploring the potential of gen AI. In December 2024, MAS published an information paper titled "AI Model Risk Management Report". Its purpose was to set out the good practices for AI model risk management that were observed, including those related to governance and oversight, key risk management systems and processes, and development and deployment of AI.

MAS highlights 3 key concerns and practices with implementation of gen AI in financial institutions:

1. Unique Risks in Generative AI adoption
2. Industry Approaches to Responsible AI Implementation
3. Key Risk Management and Technical Controls

Let's dive into each of these points, and consider how gen AI's evolution in the past few months impact these concerns.

1. Unique Risks in Generative AI adoption.

MAS highlights several risks in gen AI adoption:

• Higher Uncertainty & Hallucinations: Generative AI models exhibit unpredictable behaviors, making them less stable for high-risk or customer-facing applications.
• Challenges in Testing & Evaluation: The lack of established ground truth for certain use cases (e.g., summarization, text generation) complicates model validation.
• Transparency Gaps with Third-Party AI Providers: Many financial institutions deploy pre-trained models from external vendors, limiting access to training data and evaluation methodologies.
• Explainability & Fairness Concerns: A lack of established methods to explain Generative AI’s outputs raises challenges in ensuring fairness and bias mitigation.

My assessment:
Gen AI has evolved rapidly within the past few months such that some of these concerns can be addressed effectively. Here is the ground truth as of late February 2025:

• While higher uncertainty and hallucinations do occur in gen AI models, there are several ways that we can reduce these risks:
  • To start with, we can set the temperature of models to be 0. While this reduces the creativity of models, they are more likely to follow instructions and their provided context. As a result, hallucination is reduced.
  • Retrieval Augmented Generation (RAG) or very long context length (Gemini 2 Pro has a context window of 2 million tokens) can help to ground the model to retrieve grounded truth.
• While model validation may be more complicated than traditional supervised AI, the same principles in evaluating supervised AI models can be applied to gen AI. We still need to create eval sets to evaluate the performance of our models, and we can use LLM as a judge with human in the loop to ensure that our model is performant.
• Explainability in gen AI models has drastically improved, especially with the release of reasoning models in the past few months (OpenAI o1/o3, DeepSeek r1, xAI Grok 3). These reasoning models are able to provide their chain of thought/reasoning trace, so that you can evaluate their validity.

2. Industry Approaches to Responsible AI management

• Incremental Deployment Strategy: Most banks are limiting Generative AI applications to internal operational efficiency rather than direct customer interactions to mitigate risks.
• AI Governance & Experimentation Frameworks: Institutions are rolling out pilot programs with clear risk-mitigation controls, time-limited deployments, and human-in-the-loop oversight.
• Cross-Functional Risk Controls: Organizations are embedding multi-stakeholder risk assessment processes across development, validation, and deployment phases.

My assessment:

• One axis to deploy gen AI is internal vs external facing. As with any new tool, trust has to be built over time. By deploying gen AI for internal facing tasks first, management and employees can build trust in this system, and catch errors before they are exposed to the clients/customers.
• Internal deployment also aids human-in-the-loop monitoring, where employees are able to understand how these models function with different inputs. This can give employees the confidence to subsequently create their own eval set, which can then be used to more accurately evaluate the performance of these gen AI applications in the future.

3. Key Risk Management & Technical Controls

• Comprehensive AI Model Evaluation: Leading banks conduct multi-layered testing frameworks, including functional, standalone, and end-to-end performance assessments.
• Robust Guardrails for Input & Output Filtering: Automated filters analyze AI-generated outputs to detect issues like toxicity, bias, and privacy-sensitive information leakage.
• Secure & Compliant Model Hosting: Some institutions leverage private cloud or on-premise hosting to contain sensitive data and meet regulatory expectations.
• Retrieval-Augmented Generation (RAG) for Credibility: Institutions are experimenting with grounding AI outputs using internal knowledge bases to improve accuracy and provide source citations.

My assessment:

• Anthropic recently released Constitutional Classifiers in February 2025, a system to prevent universal jailbreaks. While time will tell the effectiveness, research into alignment is ongoing and financial institutions can leverage these understanding to improve the security of their gen AI deployment.
• With more performant open weights model available for private deployment (Alibaba Qwen, DeepSeek, Meta Llama), secure and compliant model hosting is becoming an easier task. However, the high cost of local deployment through purchase of GPUs remain a potential roadblock for enterprises who are less certain of the ROI.
• As mentioned earlier, the increase in context window of AI models can also help the models to produce more grounded outputs while avoiding the limitations of RAG (e.g. inability for the model to find relevant information in the RAG embedding).

Key use cases of Gen AI in banks

McKinsey Global Institute's 2024 estimates project that AI could contribute up to US$340 billion (S$450.7 billion) annually to the global banking sector. Thus, it's not a surprise that many banks have already piloted the usage of gen AI in their workflow. In this section, we will examine how 3 banks in Singapore have experimented with gen AI. These banks are DBS, UOB and OCBC.

DBS

As a regional digital leader, DBS Bank has strategically embraced AI for over a decade to remain at the forefront of banking innovation. Its AI-driven transformation highlights Singapore’s supportive regulatory environment, skilled talent pool, and strong R&D investments, making it a model for financial institutions to follow.

Key AI Strategies and Implementation at DBS

1. Structured Approach to AI Adoption

DBS employs a purpose-driven AI adoption framework that aligns with its broader business objectives:

• PURE Framework (since 2018) ensures AI use cases adhere to ethical guidelines—Purposeful, Unsurprising, Respectful, Explainable.
• Data Chapter (2023): A cross-functional network of 700+ data professionals embedded across the bank to industrialize AI adoption.
• Employee AI Upskilling: Over 9,000 employees trained in AI and data literacy since 2021.

2. AI Use Cases & Impact Metrics

Scaling AI Beyond DBS: Spark GenAI for SMEs

DBS is extending its AI leadership beyond banking by spearheading the Spark GenAI program (launched Nov 2024) in collaboration with Enterprise Singapore (EnterpriseSG) and the Infocomm Media Development Authority (IMDA).

Objectives:

• Targeting 50,000 local SMEs in two years to boost AI adoption.
• Aligns with Singapore's Digital Enterprise Blueprint for nationwide digital transformation.
  Four-Pillar Framework to Support SMEs:

1. Discover – AI awareness through online resources and workshops.
2. Identify – Tailored AI solution recommendations for productivity & automation.
3. Adopt – Access to grants from IMDA & EnterpriseSG lowering financial barriers.
4. Secure – Cyber resilience training & cyber insurance for risk mitigation.

Senior executives emphasize the urgent need for SME AI adoption to ensure regional competitiveness in the S$352 billion digital economy of Southeast Asia by 2024.

Key Takeaways: DBS as a Model for AI-Driven Transformation

DBS exemplifies a scalable, structured, and ethical approach to AI adoption that other financial institutions and enterprises can replicate.

• Robust governance (PURE framework) ensures responsible AI deployment.
• Investing in talent & education fosters internal AI expertise.
• AI-driven operational efficiencies create significant cost savings (S$370M in 2023).
• Industry collaboration (Spark GenAI) extends AI benefits beyond banking, helping SMEs remain competitive in the digital economy.

UOB

UOB was the first Singaporean bank to pilot Microsoft 365 Copilot in frontline and back-end functions. The trial included 300 employees across branches, customer service, technology, and operations, focusing on productivity, accessibility, and collaboration.

Use Cases Across Banking Operations

• Customer service enhancement: AI-generated responses ensure clear, concise communication, improving resolution times and customer satisfaction.
• Banker productivity: AI assists in crafting personalized messages, presentations, and reports, streamlining workflows.
• Content simplification: AI refines language, removing jargon to enhance clarity and avoid misinterpretation.

Workforce Strategy: AI as a Productivity Partner, not a Human Replacement

UOB believes that AI will augment, not replace, employees. This highlights the importance of fostering trust in the technology, both from the clients and the employees.

Research in 2023 further supports this point:

• ASEAN Consumer Sentiment Study 2023: Singaporean customers prefer digital adoption for simple banking but retain a strong preference for human-assisted interaction in complex financial matters such as loans and insurance.
• Microsoft Work Trend Index 2023: 70% of employees are open to using AI to offload manual tasks, emphasizing a shift toward AI-human collaboration.

OCBC

OCBC has positioned itself as a frontrunner in the financial sector’s AI evolution, launching key initiatives that integrate GenAI into both internal operations and customer-facing services. Since 2022, the bank has progressively expanded its AI strategy, culminating in the launch of OCBC GPT (formerly OCBC ChatGPT) in November 2023.

OCBC GPT

This proprietary chatbot, built on Microsoft Azure OpenAI, has been made available to 30,000 employees across 19 countries after an extensive six-month trial involving 1,000 employees. Early results showed task completion times were reduced by up to 50%, highlighting the efficiency gains from AI augmentation.

Key Gen AI Applications at OCBC

OCBC has integrated generative AI across multiple facets of its business, leveraging AI-powered solutions for both internal knowledge management and customer engagements:

A. Employee Productivity & Knowledge Management

1. OCBC GPT (Internal AI Chatbot)
   • Assists employees with writing, research, and ideation.
   • Has answered over 1 million prompts as of May 2024.
   • Runs on a secure private cloud environment to ensure data confidentiality.
   • Accompanied by structured prompt engineering training for employees (4,000+ trained).
2. Buddy
   • AI-powered tool specialised in retrieving internal knowledge, processing 150,000 pages of internal documents and summarising meeting records.
3. Wingman
   • Supports OCBC’s software engineering teams by generating and assisting in writing code.
4. Whisper
   • AI-driven transcription tool that processes customer service calls, summarises conversations, and extracts key insights.
5. Document AI
   • Summarises financial reports and legal documents, aiding compliance and legal teams.
   • Planned upgrades will include foreign language translation and legal document management.

B. Customer Engagement & Personalization

1. AI-Powered Customer Nudging
   • Provides actionable insights to customers, such as:
     • Identifying idle cash and recommending high-interest savings accounts.
     • Analysing subscriptions and suggesting cost-reduction strategies (e.g., duplicate music streaming services).
2. Fraud Detection & Risk Management
   • AI analyses daily transactions to detect fraudulent activities and suspicious financial behaviors.
   • Enhances decision-making in loan approvals and compliance risk monitoring.
3. Automated Investment Recommendations
   • AI analyses market trends and customer profiles to provide personalised stock recommendations.

Scaling Towards the Future: AI-Driven Decision-Making

OCBC’s AI strategy is ambitious. As of 2024, its AI models make over 4 million decisions per day, spanning risk assessment, customer service, and sales automation. By 2025, this number is expected to surge to 10 million daily decisions, further embedding AI into mission-critical operations.

Strategic Considerations

Gen AI is fundamentally a paradigm shift, and this is what makes it so challenging for many enterprises to adopt it. Here are 3 strategic considerations to keep in mind while deploying your own gen AI implementation:

1. Regulatory and ethical considerations. While these considerations may seem insurmountable, the key is to break down these considerations into bite-sized pieces. Thankfully MAS has already highlighted common concerns in their report, so your focus can be on mitigating them. To start with, you want to collect as much telemetry of your gen AI application as possible. This will help by allowing you to monitor the data to spot biases and errors, and you can iterate to solve these limitations. Remember- no system is perfect from day one. The key is to have a clear overview of potential problems by observing your data.
2. Operational risks. The same process in which you collect and monitor data can be applied to operational risks. Another axis to mitigate this risk is to start with internal facing operations, similar to how OCBC implemented OCBC GPT.
3. Talent and infrastructure gaps. Employees may be resistant to change as they are uncertain about this new technology. Setting a trial like what UOB did for their employees can encourage your employees to try out the technology. In addition, a recent McKinsey study released in January 2025 found that millennials are the most proficient age group at using these AI tools. Leverage this by creating study groups for teams in the same functions where proficient AI users can share their learnings and tips for others.

The Way Forward: Recommendations for Financial Institutions

With the rapid changes in gen AI, it can be easy to be in flight mode and not adapt fast enough. Here are 3 tips to help your enterprise adopt gen AI:

1. Embrace "human in the loop". Gen AI is still better used as the co-pilot in most enterprise use cases. Start with augmenting human expertise, not replace it. This will build trust in the AI applications, and help to highlight any potential problems.
2. Start with high-impact, low risk AI applications. Use AI in internal applications first- this can help your employees familiarise themselves with this technology, and allow management to build trust.
3. Partner with regulators, vendors and ecosystem players. The rapid advancement in AI means that the best vendor for you today may not be the best vendor for you tomorrow. Ideally you will want to partner with consultants who keep a pulse on the whole AI industry so that you know what is the most appropriate tool for you. This will allow you to focus on your customers.

Conclusion and Call to Action

Gen AI is reshaping finance- don't get left behind. Assess your workflows and explore AI-driven efficiencies. Need expert guidance? Let's connect.