AI in banking: The good, the bad and the ugly

ARTIFICIAL intelligence (AI) and machine learning (ML) are widely celebrated as the most influential technology and business trends today.

McKinsey research suggests that a convergence of 15 disruptive technologies – from cloud to climate – will transform the future of businesses and organisations. Some US$1 trillion in capital has been invested into companies producing or linked to these.

Of the 15 technologies, three have witnessed high maturity in terms of business adoption – applied AI, cloud computing and advanced connectivity. Generative AI (GenAI), the newest kid on the block, may be low on the maturity scale but also holds enormous potential for businesses, especially financial institutions and other service industries.

For these reasons, even the harshest critics say there is so much potential for AI/ML. Speaking as a tech specialist in banking, I can say that these technologies have truly created value in financial institutions. Yet, I am also concerned about the risks involved.

The good: How AI creates value in banking

While the explosion of available data and sophistication of statistical techniques may be a topic of fascination for data scientists and statisticians, the real question for practitioners is whether AI/ML really drives business value. The answer is a resounding “yes”.

Leading financial institutions and fintechs are using AI/ML to drive insights and granular decisions across the customer life cycle from acquisition, credit decisioning, servicing, cross-selling and retention.

This requires hundreds of ML models across multiple domains, including credit analytics (customer scoring models, credit pre-approval models to drive lending decisions); marketing and personalisation (product and channel propensity models to drive personalised recommendations for customers); engagement and retention models (attrition models to identify propensity and reasons to attrite); and a full suite of risk and collections models (loss forecasting models, pre-delinquency management to optimise customer repayments).

AI/ML is also used to improve the efficiency of internal processes and lift productivity of employees. For example, OCBC uses AI to augment high-volume processes, such as anti-money laundering reviews and scam detection. AI is used to automatically identify potential anomalies and close alerts that are deemed as false positives without human investigators being involved. This can reduce investigator effort by as much as 70 per cent.

AI also plays a role in enhancing the customer experience. OCBC leverages hyper-personalisation to improve the relevance of our customer communications across channels.

Over 100 million personalised offers are displayed to customers to show them the most relevant “Next Best Product”, relevant credit card deals or the investment research articles most appropriate to their profile. Our share-trading subsidiary, OCBC Securities, also uses “Oscar” – an AI-powered engine which makes intelligent stock recommendations to our customers.

Finally, AI is used to monitor unstructured customer feedback received across channels such as surveys, complaints, social media and the various app stores. The algorithms analyse and categorise the huge volume of feedback, providing the business with actionable insights on emerging servicing issues.

Now, how does the arrival of the latest natural language processing (NLP) or GenAI model impact banking? Can we envision a future bank where all customer interactions, including sales, portfolio advisory, transactions and servicing be driven by GenAI-powered machines? While the use of GenAI in banking is nascent, there are at least four areas of possibilities that are emerging:

1) Operations automation

An example of this is customer-focused chatbots, which streamline customer queries. One report estimated that operational cost savings from effectively using chatbots in banking could reach US$7.3 billion globally, equivalent to roughly 862 million hours of time saved.

Our NLP-powered chatbot in OCBC already handles around one in three customer service enquiries. Powered with GenAI, there is potential for chatbots to take on an even larger role in servicing and delighting customers.

2) Virtual experts

Voice-enabled bots interact with customers and play the role of the traditional relationship manager by addressing customer queries and service transaction requests. They can provide intelligent insights and recommendations on customer portfolios, deferring to humans only for highly complex queries or in events that have legal or tax-related implications.

As an illustration, while a virtual expert can provide recommendations on high-performing funds to invest in, it cannot provide advice on tax implications of structuring portfolios.

3) Content generation

In-house marketing teams can now leverage the power of GenAI’s ability to generate text, images and videos to create personalised marketing content directed at customers. Coupled with ML model outputs on the “Next Best Action” for a retail bank customer, GenAI can help define “how” to communicate the offer to the customer by designing appropriate content, copy and creatives that are tailored to individual customers. This was a task that, up until now, was purely restricted to creative geniuses at marketing agencies.

This capability, when integrated with marketing automation technologies (dynamic content optimisation, campaign automation, et cetera) transform precision marketing at scale. Indeed, many platforms are already experimenting with native integrations of GenAI within their platforms.

4) Code acceleration

Programming copilots such as GitHub Copilot or our own internal “OCBC Wingman” are improving the productivity of developers and engineers by “writing” a first version of software code and automating tasks, such as debugging, refactoring and documentation.

As this matures, the role of the programmer will shift from authoring code to checking and refining code with dramatically improved productivity. Market and internal OCBC tests suggest a productivity improvement of 20 per cent imminently achievable.

The bad and the ugly: The dark side of AI

As GenAI-based applications increasingly move from the research lab to business, executives and policymakers are debating the “dark” side of AI.

Banks do have a lot at stake, and it is imperative for them to be fully cognisant of the risks involved in deploying AI/ML/GenAI to drive bank operations, services and transactions. They must work to protect their customers’ data, retain trust with them and so, maintain institutional reputation.

While there are established guidelines and principles, such as Feat, that promote fairness, ethics, accountability and transparency in the use of AI and data analytics in Singapore’s financial sector, there are at least four risks that banks need to protect against:

• Ethical and reputational risks: GenAI algorithms that are trained on data available on the Internet reflect the inherent bias in social media. Banks need to be extremely cautious when using GenAI-generated content (for example, text, images and videos) to ensure reputational and/or legal risks. Having a human review content is, of course, possible. But how would this be possible if a bank uses tools to generate thousands (possibly millions) of pieces of personalised content for individual customers? 
• Promoting inequality: Do AI/ML-driven algorithms promote inequality in a bank’s customer base? Risk and collections models that are routinely used by banks may detect underlying risky behavioural trends that are concentrated, say, in a particular geography or in a particular demographic group. If 60 per cent of customers in a particular micro-market default on their instalment payments, is it “fair” to assume that all customers in that geography are “risky”? We must ensure that bias in underlying data is understood, so that we do not perpetuate these in future model recommendations and cause unnecessary harm to sensitive customer groups. 
• Transparency: When using AI/ML in a material process such as credit decisioning or money laundering, banks need to have a high degree of understanding of how the model reached its decisions. Due to the large, complex nature of GenAI models, being able to explain how they produce an output is very challenging. This explainability risk is a barrier that will need to be overcome before widespread adoption of GenAI is possible within critical banking processes. 
• Data security and manipulation: Data and AI/ML are immensely powerful in the right hands. In the wrong hands, it can be manipulated to enable unethical or criminal activity, such as fraud. Banks must think through sufficient layers of protection of their data and models against manipulation by external agents and/or by internal employees. 

Finally, if we envision a future bank powered by AI algorithms, one has to question what this means for a bank’s employees. To me, the answer lies in continuing to complement and augment employees to be AI-ready, which includes the need to upskill workforces with the introduction of new AI and its technology capabilities. Only then can banks be more effective, productive and deliver an even greater experience to customers.

The writer is head of group operations and technology at OCBC