Signals | Generative AI: transforming banking

Signals

Q4 2023

Not long ago, in the fall of 2022, generative AI was primarily known to AI engineers and data scientists. Today, this technology has garnered widespread awareness and stands at the forefront of an economic revolution. With 55% of the CEOs of large global companies surveyed in 2023 indicating that they are “evaluating or experimenting” with gen AI and 37% that they are already using it, the technology has the potential to unlock trillions of dollars in global economic value.

Mastercard first explored gen AI's implications in Commerce in the age of generative AI, our Q3 2023 Signals issue. Since then, we've worked with the technology while carefully observing and evaluating its critical use cases in financial services. In areas like software development, it has the potential to be transformational. It’s also a useful tool to consolidate structured and unstructured data for customer service teams and has promising applications in cyber protection and in creating more personalized experiences and services. Yet hurdles stand in the path to an AI-enabled banking future: data privacy concerns, the potential for bias and the proliferation of disinformation, to name a few. As the industry grapples with these issues, many early adopters are deploying gen AI internally to improve operations. The rate at which the tech becomes commonly adopted should hinge on the industry's ability to ensure the accuracy of outputs, integrate safeguards and ethical standards, and comply with global regulations. In this edition of Signals, we look at some of the promising gen AI use cases that could reshape banking. We unpack the challenges and potential solutions to mitigate them and chart the paths banks may take to adoption.

The AI-powered bank

Numerous use cases are emerging to achieve efficiencies, accelerate productivity and reimagine customer experiences.

Mitigating challenges

Generative AI technology comes with unique challenges, making it critical that banks confront the ethical and regulatory issues it presents.

Data, the lifeblood of generative AI, is a resource banks must handle with sensitivity. Keeping client data private isn't only a regulatory mandate, it’s a business imperative. Banking is built on trust.

Data privacy: Ensuring ethical innovation

Inaccuracy: Navigating the mirages of AI

Being responsible stewards of data takes on an additional importance in the broader AI context, including gen AI. This includes ensuring data integrity, the maintenance of information that's accurate, credible and free from errors or biases. It also requires that data and algorithmic system outputs be traceable, explainable and trustworthy. The challenges lie in governing unstructured data, which has not previously been used in building LLMs, and in illuminating the "black box" of gen AI: neural networks’ inner workings are often as enigmatic as they are complex, making it hard to discern the "why" behind an AI's decision. In addition, data responsibility requires providing models with data that encourages them to produce equitable and accurate outputs. A bias within an AI often reflects a bias in its training data: An LLM trained on historical data reflecting discriminatory lending practices could perpetuate or even amplify such practices.

The data needed by gen AI models often comes with constraints. Technological barriers at banks can strand data within inaccessible systems and storage environments.

Data availability: Managing access and sovereignty

Regulatory uncertainties

As gen AI weaves deeper into the fabric of global finance, it operates in an environment that regulation still needs to define fully.

United States: An executive order

President Biden's October 30, 2023 Executive Order represents the U.S. government's most recent move toward regulating AI. The order calls for a range of measures to boost safety and security, ensure privacy, advance equity and civil rights, protect consumers and patients, support the workforce, encourage innovation, promote U.S. leadership in AI and foster "responsible and effective" government deployment of the tech. In addition, legislative activity on AI is ongoing at all levels of government.

Governments worldwide have been playing legislative and regulatory catch-up with gen AI, grasping at it as it takes off. The result for now is a global kaleidoscope of regulatory approaches that reflect the priorities and strategies of the countries that have instituted or are planning to institute them. Gen AI legislation also exists within the context of more broadly applicable regulations governing data usage and privacy.

Gen AI poses complex questions around data stewardship for an industry responsible for clients’ most sensitive information. How can banks protect information while using AI to enhance service delivery? What measures can ensure data integrity in light of cyber threats and misinformation? Can AI systems serve all customers equitably, free from prejudices that may impact human decision-making? These questions are not merely rhetorical. Banks bear the responsibility of maintaining trust and ensuring compliance in an AI-augmented future. It's a delicate balance — cultivating AI innovation with one hand while protecting against potential misuse and unintended consequences with the other.

Gen AI can help improve essential banking functions by taking on many data-heavy tasks. Internally, banks can unlock hidden insights, forecast with better precision and augment staff productivity. Externally, the era of one-size-fits-all banking could give way to a new age of individualized engagement, where gen AI facilitates personalized experiences.

Alongside infrastructure development, progress is likely contingent upon a growing reservoir of AI expertise within the banking sector. Difficulties in assembling the talent, hardware and capital required to build proprietary LLMs could limit such ventures to larger institutions. The AI talent crunch in particular is by now well-documented. Some 75% of companies for whom hiring AI specialists is a priority reported being unable to fill their AI talent requirements, according to November 2023 research. These difficulties could continue to power the developing market for specialist intermediary companies that build, host and support gen AI capabilities for organizations that lack sufficient capabilities of their own.

Technology

Cybersecurity and fraud prevention

One of the critical capabilities of gen AI is the ability to synthesize vast amounts of data into actionable form. Bankers equipped with gen AI may find that information searches that once consumed hours could now take minutes or seconds. When they need to check up on details of complex regulations, bankers could, via gen AI, receive cogent text or voice summaries — rather than just citations of, or links to, statutes and other raw material, often written in legal jargon.

Talent management

Client onboarding

Credit issuance

Conversational banking

Wealth advisory

Charting a course through the fog

For now, banks and other financial institutions are confronting the inaccuracy challenge by taking a conservative approach to gen AI. Early applications have been internal: Banks mainly use it to power internal knowledge management and research solutions.

A blueprint for better accuracy

To uphold data integrity, banks will draw on tried-and-true methods. They’ll also undertake innovative practices tailored to the nuances of gen AI:

The quest for open data

Anticipated technological advancements promise to dismantle specific barriers that impede data flow. Banks can look forward to the following developments:

The growth of APIs is set to enhance connectivity, creating data-sharing conduits that ease access and improve utility.

Expanding API networks

An advancing plug-in infrastructure will let generative AI models act as gateways, with users tapping into integrated services seamlessly through them.

Plug-in proliferation

Crucially, data availability must be viewed in tandem with data-related regulations. An expanding legal and regulatory web and jurisdictional realities can limit access due to privacy concerns, data localization mandates and evolving concepts of data sovereignty.

The EU’s Artificial Intelligence Act, which could be finalized by the end of 2023, is meant to balance AI's benefits against the need to safeguard the public interest. The risk-based legislation, development of which started before the November 2022 launch of ChatGPT, would consider AI use for credit scoring as high risk unless it poses no significant harm to individuals. It introduces transparency requirements for generative AI as well. Late October 2023 saw EU legislators progress toward laying down criteria for "high-risk" AI systems, long debated in the law's development.

European Union: Top-down bloc-wide law

The U.K. is charting a different path, eschewing broad new legislation on the assumption that its existing legal frameworks will be sufficient for the AI revolution. This regulatory approach aims to promote innovation by providing bespoke guidance in discrete, sector-specific situations, including via an "advice service" staffed by representatives of relevant government agencies. This service would provide "tailored support to businesses so they can meet requirements across various sectors while safely innovating" in gen AI. The service is set to launch in 2024.

United Kingdom: A pro-innovation approach

India is formulating new policy frameworks as regulators balance innovation and risk mitigation. Its regulatory roadmap focuses on capacity-building, digital infrastructure and support for creation. India also indicated that it would establish ethical guidelines for gen AI developers.

India: Formulating a policy

Like the EU’s, Brazil’s AI legislation concentrates on protecting the individual. It establishes the rights of consumers and mandates that providers of AI solutions be able to explain how their technologies arrived at specific outputs. Consumers are also entitled to know when they’re communicating with an AI. Again like the EU’s, the law classifies AI solutions into risk categories, with providers of higher-risk solutions facing relatively more liability for potential harm. AI tools deemed of the highest risk levels are restricted.

Brazil: Focusing on user protection

China has been an early mover in AI regulations, striving to balance national security with global competitiveness. It's taken a targeted and iterative approach to boosting innovation, with regulatory measures that require security assessments only from companies creating gen AI-driven products for consumers, exempting creators of enterprise tech. In October, China's government announced security-related requirements for gen AI firms to follow. Pools of training data in which at least 5% of the information is "illegal and harmful" would be banned.

China: Balancing security and innovation

From exploration to deployment

The inevitable integration of gen AI into banking will likely happen in a phased manner — and be subject to a range of considerations.

The use cases we discuss in this Signals issue are achievable using LLMs available today in tandem with other emerging methods, technologies and assets that enrich gen AI capabilities, like open banking, IoT and blockchain data. The move from exploration to full-scale deployment of generative AI will likely unfold in calculated, incremental steps.

Fueling the journey — talent

Blueprint for a gradual rollout

Progress toward a future significantly defined by gen AI will come in phases:

An interesting phenomenon of the first year of "mainstream" gen AI has been how leading AI developers have explicitly called for more government regulation of their industry. Some industry observers believe there could be an element of protecting first-mover advantage in this, as gen AI’s current winners attempt to sideline the competition. But they also know that legal, regulatory and even ethical clarity from regulators could reduce liability and potential missteps that may have consequences for their clients and themselves.

Regulatory guardrails for safe passage

The astounding pace at which generative AI has advanced prompts speculation about a potential plateau in computational power for industry leaders like OpenAI, which launched GPT-4 in March 2023. GPT-4 is estimated to contain more than a trillion parameters (an indicator of LLM power), while GPT-2, released in winter 2019, deployed a mere 1.5 billion parameters. The law of diminishing returns could be at play here, as gen AI users can get the functionality they need from older, smaller LLMs, without paying a premium for newer, bigger ones. Still, AI companies other than the marquee-name giants have headroom to improve their LLMs significantly, widening the ecosystem and the range of available options. If stabilization occurs, it may mark the beginning of a period of integration, when banks consolidate what they’ve learned and focus on harnessing the capabilities of current gen AI models rather than waiting for further innovations.

Navigating a potential plateau

Careful experimentation, clear use cases and strategic caution characterize gen AI in banking today. As trust in its capabilities grows, generative AI is poised to become an integral thread in the fabric of financial services. The adoption of generative AI could reshape the competitive dynamics within banking, potentially empowering new market entrants or altering the relative influence of existing institutions. The stage is set for a deeper exploration of gen AI strategies, challenges and potential, which could not just augment the banking experience for consumers but transform it fundamentally.

Today's innovations, tomorrow's foundations

Generative AI, as formidable as it is in predicting and producing data sequences, can be subject to "hallucinations" — instances where it conjures up plausible but incorrect, nonsensical or irrelevant outputs — and inaccuracy in general. Poor training data, model complexity and imprecisely worded user prompts exacerbate these liabilities, which can be mitigated but not eliminated. The finance sector is especially vulnerable to such challenges because false data could mislead investors or, in extreme cases, shock the economic system.

Loyalty programs

Marketing and communications

Navigating the AI policy patchwork

As they navigate this new digital landscape, banks could adopt multiple strategies to mitigate misinformation and accuracy challenges:

Other measures for handling gen AI's hallucinatory tendencies and other accuracy issues will likely emerge given the rapid evolution in this field.

At Goldman Sachs, generative AI is at the heart of several proof-of-concept initiatives. These pioneering efforts envision using the technology for document categorization, precise information extraction and sophisticated coding tasks.

Knowledge and insights

Onboarding, especially for corporate clients, tends to be time- and resource-intensive. Research indicates that onboarding can take up to 100 days, a significant delay before a bank starts generating revenue from a product or service it has sold. Integrating gen AI into bank client onboarding processes promises improvements in both bank efficiency and customer experience. Client onboarding managers could streamline know-your-customer, or KYC, compliance and documentation management, which are substantial components of the onboarding process. Rapidly synthesizing client data, it could flag risks more efficiently than traditional methods. It may also automate the completion and organization of paperwork, expediting timelines. On the customer end, generative AI bots could interact intelligently with applicants to resolve ambiguities in the application and onboarding processes, reducing the need for back-and-forth communication with bank staff. Banks might improve their time-to-revenue metrics by reducing the time from initial engagement to the start of a financial relationship.

Client onboarding

of the bank client onboarding process is taken up by KYC and account-opening tasks.

Data privacy: In the U.S., regulations such as California's Consumer Privacy Act are important in the absence of federal legislation dictating terms of data management compliance. Biden’s October 30 Executive Order, with its stipulations regarding privacy, is also relevant here.

Data privacy: The EU's landmark Global Data Protection Regulation, which took effect in 2018, prescribes stringent data protection and privacy measures, setting a high standard for protecting personal data.

Data privacy: Chinese regulations delineate a unique set of data governance rules influenced by national policies and security concerns.

At Mastercard, AI models already serve as linchpins to our solutions, safeguarding more than 125 billion transactions on our network every year. Employing hundreds of data scientists, AI technologists, and a growing team of AI governance experts, we’re committed to developing practical AI solutions that integrate privacy and ethics by design. Across our capabilities — data intelligence, open banking, identity, fraud protection and cybersecurity — Mastercard ensures trust is at the forefront, and AI is used responsibly and ethically.

Coming next in Signals: Emerging tech trends for 2024

Produced by Mastercard Foundry

View sources

Sources 1. Deloitte CFO Signals: 3Q2023 2. The Wall Street Journal: Goldman Sachs CIO Tests Generative AI 3. OCBC: OCBC is first Singapore bank to roll out generative AI chatbot to all employees globally 4. InfoWorld: IBM Watsonx to use generative AI to translate COBOL code into Java 5. Diginomica: Deutsche Bank doubles down on generative AI after laying foundations with Google Cloud 6. Google: Supercharging Security with generative AI 7. McKinsey: Winning corporate clients with great onboarding 8. McKinsey: Winning corporate clients with great onboarding 9. Business Wire: Kasisto Launches KAI-GPT, the First Banking Industry-Specific Large Language Model 10. McKinsey: Reimagining customer engagement for the AI bank of the future 11. CNBC: Morgan Stanley kicks off generative AI era on Wall Street with assistant for financial advisors 12. Yahoo!Finance: Meet ‘IndexGPT,’ the A.I. stock picker JPMorgan is developing that may put your ‘financial advisor our of business’ 13. Bloomberg: JPMorgan Is Discussing Its Generative AI Projects With Regulators 14. Finicity: Open banking drives new mortgage capabilities 15. Payments Journal: Loyalty Programs Are Great — But What About “Everyone Else”? 16. Statista: Loyalty programs and marketing in the U.S. — statistics and facts 17. Salesforce: Top Generative AI Statistics for 2023 18. Persado: Credit Agricole personalizes customer communications in an omni-channel environment 19. IBM Newsroom: NatWest and IBM Collaborate on Generative AI Initiative to Enhance Customer Experience 20. SouthState Correspondent Division: Generative AI — 7 Lessons That Tate Taught Us 21. CNBC: Bloomberg plans to integrate GPT-style A.I. into its terminal 22. The Economic Times: Big banks take to large language models trained on internal data 23. The New York Times: Chatbots May Hallucinate More Often Than Many Realize 24. FACT SHEET: President Biden Issues Executive Order on Safe, Secure, and Trustworthy Artificial Intelligence 25. The AI Act: Documents 26. Human Rights Watch: Artificial Intelligence Regulation Should Ban Social Scoring 27. Key Issues: Transparency Obligations 28. Reuters: EU lawmakers make progress in crucial talks on new AI rules — sources 29. GOV.UK: New advisory service to help business launch AI and digital innovations 30. TechCrunch: India opts against AI regulation 31. DigWatch: Indian government is not considering any law to regulate AI 32. The Washington Post: From China to Brazil, here’s how AI is regulated around the world 33. T_HQ: China finalizes rules on generative AI amidst a rush of product launches 34. Reuters: China proposes blacklist of training data for generative AI models 35. Medium: Doing it right with AI: Ally’s foray into generative AI 36. Investopedia: Amazon Launches AI Training Program As Companies Contend With AI Talent Shortage 37. The New York Times: OpenAI’s Sam Altman Urges A.I. Regulation in Senate Hearing 38. AP: OpenAI boss ‘heartened’ by talks with world leaders over will to contain AI risks 39. The Decoder: GPT-4 has more than a trillion parameters 40. Hugging Face: OpenAI GPT2 41. PYMNTS: Ant Group Launches Specialized Large Language Model for Finance Industry

Hide sources

Generative AI: The transformation of banking

Hallucination rates of leading gen AI solutions:

A prerequisite for this expanded data access will be for banks to possess the mechanisms and infrastructure to share data with LLMs. That includes consent-based mechanisms, like those used in open banking ecosystems: Users of an app will be informed that the data they supply will feed into an LLM and will have to give their permission for this. In addition, banks must have the digital foundations and infrastructure to tokenize data.

Signal: Ally Financial launched its Ally.ai platform with a solution providing real-time transcription and summing up client help-desk phone conversations. Ally reported that over 80% of the AI summaries needed no human editing.

Signal: Chinese fintech Ant Group has created its own LLM. It will power, among other products, an assistant for financial industry professionals and an application that answers customer queries.

The technology could change how banks deliver and customers interact with financial services through three key capabilities:

View sources

Sources 1. The Majority of CEOs Believe Generative AI will Increase Their Organizations’ Efficiencies: ‘Summer 2023 Fortune/Deloitte CEO Survey’ 2. The Wall Street Journal: Goldman Sachs CIO Tests Generative AI 3. OCBC: OCBC is first Singapore bank to roll out generative AI chatbot to all employees globally 4. Generative AI in Software Development 5. How Generative AI improves the productivity of Software developers 6. How generative AI correlates IT and business objectives to maximize business outcomes 7. InfoWorld: IBM Watsonx to use generative AI to translate COBOL code into Java 8 Diginomica: Deutsche Bank doubles down on generative AI after laying foundations with Google Cloud 9. Google: Supercharging Security with generative AI 10. McKinsey: Winning corporate clients with great onboarding 11 McKinsey: Winning corporate clients with great onboarding 12. Business Wire: Kasisto Launches KAI-GPT, the First Banking Industry-Specific Large Language Model 13 McKinsey: Reimagining customer engagement for the AI bank of the future 14. CNBC: Morgan Stanley kicks off generative AI era on Wall Street with assistant for financial advisors 15. Yahoo!Finance: Meet ‘IndexGPT,’ the A.I. stock picker JPMorgan is developing that may put your ‘financial advisor our of business’ 16. Bloomberg: JPMorgan Is Discussing Its Generative AI Projects With Regulators 17. Finicity: Open banking drives new mortgage capabilities 18. Payments Journal: Loyalty Programs Are Great — But What About “Everyone Else”? 19. Statista: Loyalty programs and marketing in the U.S. — statistics and facts 20. Salesforce: Top Generative AI Statistics for 2023 21. Persado: Credit Agricole personalizes customer communications in an omni-channel environment 22. IBM Newsroom: NatWest and IBM Collaborate on Generative AI Initiative to Enhance Customer Experience 23. SouthState Correspondent Division: Generative AI — 7 Lessons That Tate Taught Us 24. CNBC: Bloomberg plans to integrate GPT-style A.I. into its terminal 25. The Economic Times: Big banks take to large language models trained on internal data 26. The New York Times: Chatbots May Hallucinate More Often Than Many Realize 27. FACT SHEET: President Biden Issues Executive Order on Safe, Secure, and Trustworthy Artificial Intelligence 28. The AI Act: Documents 29. Human Rights Watch: Artificial Intelligence Regulation Should Ban Social Scoring 30. Key Issues: Transparency Obligations 31. Reuters: EU lawmakers make progress in crucial talks on new AI rules — sources 32. GOV.UK: New advisory service to help business launch AI and digital innovations 33. TechCrunch: India opts against AI regulation 34. DigWatch: Indian government is not considering any law to regulate AI 35. The Washington Post: From China to Brazil, here’s how AI is regulated around the world 36. T_HQ: China finalizes rules on generative AI amidst a rush of product launches 37. Reuters: China proposes blacklist of training data for generative AI models 38. Investopedia: Amazon Launches AI Training Program As Companies Contend With AI Talent Shortage 39. Medium: Doing it right with AI: Ally’s foray into generative AI 40. The New York Times: OpenAI’s Sam Altman Urges A.I. Regulation in Senate Hearing 41. AP: OpenAI boss ‘heartened’ by talks with world leaders over will to contain AI risks 42. The Decoder: GPT-4 has more than a trillion parameters 43. Hugging Face: OpenAI GPT2 44. PYMNTS: Ant Group Launches Specialized Large Language Model for Finance Industry

Hide sources

Signal: Sam Altman, CEO of ChatGPT creator OpenAI, is among the high-profile tech leaders calling for gen AI regulation. Altman has met with over 100 U.S. legislators and several world leaders to make his case.

As regulation matures, it will guide banks through uncharted territory. It will provide the assurance that moving forward with gen AI is both ethical and compliant, fostering a secure environment for innovation. For more on ethical data practices, please click here.

40, 41

Synthesis of data

Creation of content

Human-like engagement

“Earning trust comes down to making a promise and delivering on it again and again. It starts with building robust AI governance that focuses on privacy and security, transparency, fairness and accountability — putting people at the center of all we do, for the benefit of the whole financial ecosystem.”

Caroline Louveaux,

Mastercard’s Chief Privacy and Data Responsibility Officer

Introducing additional data can help an AI model recalibrate when it starts hallucinating, as well as make it generally more accurate for specialized use cases.

Focused model tuning

Building queries for LLMs is becoming a crucial skill. The most effective prompts are concise and concrete, steering clear of abstract language.

Precision prompts

Deliberately providing LLMs false or challenging prompts can fortify them against potential errors — a process akin to immunization.

Adversarial fortification

Limiting the length of an AI's responses can also reduce its propensity for fabrication. The shorter the answer, the less room for error.

Embedded brevity

Vetting conducted by human experts remains an indispensable defense layer.

Human oversight

Taking measures to track where, when and by whose agency data fed into an LLM can help make an AI system's results explainable.

Tracking data provenance/lineage

Selective use of closed vs. open-source models

Banks will likely avoid the unmediated use of closed-source proprietary generative AI solutions (such as Chat GPT) hosted on external servers. Connecting to such solutions "as is" via application programming interfaces or APIs, for example, without safeguards, could bring a risk of data exposure. To be clear, banks are not abandoning closed-source solutions. Instead, they are approaching them deliberately, which lets banks take advantage of the solutions' strengths while keeping themselves safe. The self-securing methods they are using include:

Building private models from scratch

Banks can also build their own LLMs, although it requires a considerable investment in time and infrastructure, such as cloud environments that are both secure and can handle gen AI’s heavy data loads.

There are other challenges to be overcome with this approach. Established banks have primarily built their systems to meet big data processing needs and are not technically ready to take on gen AI in scale. They need graphics processing units, or GPUs, the hardware crucial to gen AI development, which have been costly and in demand, to the point of shortage. Banks that manage to acquire sufficient hardware may then face the challenge of integrating it into their legacy processing environments in a way that ensures that data can feed into the bank system fast enough. Other challenges exist. One is curating enough data to build an LLM. Another is developing the software and internal expertise necessary to do these things at scale. A third is geographical. The prominent vendors of the cloud solutions required for training LLMs do not possess uniform processing power across the globe. For example, a particular provider’s capabilities may be stronger in North America than in Europe, which would hinder European banks from performing LLM projects. At the same time, data localization requirements might keep those banks from building their LLMs elsewhere.

OpenAI solutions

Anthropic’s Claude 2

Meta solutions

Google’s PaLM

The foundational practices of data governance remain invaluable — cleansing and preprocessing data, implementing data lineage and strong security to thwart data corruption, enacting governance protocols to regulate access, training staff and practicing data minimization to limit exposure.

Legacy techniques

LLMs can undergo fine-tuning with additional training data and reinforcement learning that can counter past biases and foster outputs consistent with institutional values and objectives.

Tailoring LLMs

Persistent observation, whether human or automated, of the data feeding into AI is essential.

Vigilant monitoring

People representing demographic diversity of all types will be harnessed to consult on AI practices. They will come from both inside and outside of an organization. Some of them will even be unfamiliar with technology. They may offer perspectives on an AI's data diet, opine on ethical matters, and outline expectations or judge outcomes.

The wisdom of the crowd

Ongoing evaluation of AI-driven processes will be critical to rectifying any suspect or potentially detrimental patterns that may emerge.

Continuous auditing

At OCBC Bank in Singapore, a generative AI application is revolutionizing how employees conduct research, create documents and spark innovation. Early trials have shown promising results, slashing the time required for complex tasks by 50%.

Gen AI could play a significant role throughout the software development process. Engineers could use it in the design phase to generate iterations of product user interfaces and draft project specifications; in the development phase to write, review and debug code; in the testing phase to create synthetic data with which to stress new solutions as well as to develop test-cases; and in later lifecycle stages to expedite the error-intensive work of refactoring code.

Technology

IBM’s Code Assistant for IBM Z, released in late 2023, is a gen AI-driven refactoring product that translates COBOL, the circa-1959 legacy coding language still fundamental to many banks’ IT systems, into modern Java.

Since 2020 Deutsche Bank has been transitioning its technical infrastructure to the cloud. This strategic initiative has fortuitously laid a foundation for integrating generative AI into core operations, such as software engineering and account management.

Cybersecurity professionals could also benefit from integrating generative AI into fraud detection models. Specialized large language models, or LLMs, that are either custom-built from scratch or refined versions of pre-existing models can be tailored for security tasks. They could investigate and respond to threats immediately upon detecting them and synthesize complex data into clear guidance that professionals can act on.

Cybersecurity and fraud prevention

SecPaLM from Google is an LLM trained especially for use in security applications. It will power the Google Cloud Security AI Workbench, a product that offers a range of tools, including for identifying malicious code and responding fast to breaches.

“Integrating generative AI technology allows financial institutions to identify intricate patterns, anomalies and evolving fraud strategies that traditional methods might miss, bolstering security measures and safeguarding against potential risks.”

Rohit Chauhan,

Mastercard’s Executive Vice President for Artificial Intelligence

Legacy chatbots have long been a staple in digital banking. Operating on rigid decision-tree algorithms, they often fail to meet customer expectations, struggling with the nuances of human dialogue.

Conversational banking

Startup Kasisto launched KAI-GPT, the first LLM created especially for banking, in spring 2023. It gives bank chatbots human-level context-responsive conversational abilities.

Customer satisfaction can translate into value: U.S. retail banks with top satisfaction scores saw their deposits grow more than 80% faster than their lowest-rated peers, according to McKinsey.

LLMs could take on some of the functions of a financial advisor, benefitting from an ability to synthesize vast amounts of information rapidly and to offer advice unencumbered by human emotions, preconceived notions or wishful thinking.

Wealth advisory

Morgan Stanley Wealth Management has launched an AI-powered knowledge management assistant. This assistant, which utilizes OpenAI's ChatGPT-4, helps the firm’s advisors by providing easy access to an internal repository of 100,000 documents and reports.

JPMorgan Chase is developing an AI agent, IndexGPT, that will choose securities and other financial assets. Whether the tool will be meant for the public or for the institution’s employees remains to be determined. The firm is also testing a gen AI-powered customer-help solution and one that creates earning reports for companies the firm is monitoring.

of respondents in a survey performed by Finicity, a Mastercard company, indicated that their mortgage application experience was at least as stressful as the process of finding a home itself.

Loan officers assessing credit applications from small and medium-sized businesses could find that gen AI simplifies complex application processes. Gen AI may reduce processing times and associated costs by offering applicants step-by-step conversational guidance. More transparent and straightforward applications methods could compel more of these businesses to apply for loans, spreading the benefits of credit more widely throughout the economy. Retail credit could also benefit from gen AI, with similar solutions that ease consumer lending processes, like those for home mortgages. In addition, LLMs could help improve the quality and accuracy of data that consumers share through open banking protocols, improving the consumer experience. In tandem with other AI models, generative AI could also revamp point-of-sale lending by leveraging contextually rich data about consumers’ financial needs and personal circumstances to offer more relevant options. With its advanced interactive capabilities, gen AI could hold real-time conversations with consumers, providing immediate clarifications on terms and conditions, thus demystifying financial offerings.

Credit issuance

of loyalty program users go inactive within the first 90 days of signing up.

Rewards program executives could see gen AI change how they craft, manage and enhance loyalty experiences. Gen AI gives program managers a possible tool with which to communicate with participants about their desires in real time. Its ability to synthesize customer data from diverse sources — including unstructured data sources such as social media sites, blog posts and crowdsourced review sites — may give program managers another view into what program users want. Armed with this information, managers can better match offers to users. Additionally, navigating loyalty programs can be daunting, given their often intricate options and use of industry jargon. Gen AI could streamline this process by enhancing personalization and relevance. Such a paradigm shift, fueled by gen AI agents that are responsive and proactive, could benefit all stakeholders in the loyalty ecosystem. Customers would appreciate better personalization and more meaningful offers. Companies would see an uplift in engagement and revenue.

Loyalty programs

The average American belongs to more than 16 loyalty programs but is an active participant in just half of those.

of marketers say that gen AI will relieve them of rote work, freeing them to concentrate on strategic tasks.

Generative AI is poised to transform marketing through its ability to generate a virtually limitless variety of customized content. Marketing executives could use it to perform dynamic testing and to optimize communications, such as campaign emails and social media posts. By pairing gen AI content generation with sentiment analysis and social listening tools, marketers can gain an intricate understanding of consumer reactions. Marketing strategists could find that gen AI encourages recalibration of the relative importance of marketing strategies. Product-centric campaigns, in which finished products and services are pushed out to the public through traditional bank marketing channels, could cede space to those that develop around client-centric dialogues, carried out via gen AI chatbots and other tools. These dialogues would reveal clients' financial challenges and priorities — in other words, what they need, want and would like to see on the market. The result: more pertinent product and service offerings.

Marketing and communications

Credit Agricole, France’s second-largest bank, is using gen AI solutions from AI company Persado to create marketing emails as well as Facebook and Google Ads copy.

“There’s not going to be a single LLM that will be all things to all people, even in a single organization. Every use case might have its own specific, custom LLM, a situation that will create its own infrastructure issues.”

Steve Flinter,

Mastercard's Distinguished Engineer, AI and Quantum Computing

Florida-based SouthState Bank in spring 2023 launched Tate, a gen AI-powered knowledge management bot for its employees’ use. The bot is based on OpenAI’s ChatGPT and was built in Microsoft’s Azure cloud. According to the bank, it cut info search times from an average of seven minutes to under 32 seconds.

The U.K.’s NatWest is augmenting its chatbot, Cora, with generative AI capabilities provided by IBM’s watsonx platform. The new bot, Cora+, will offer customers information about bank services and more.

HDFC Bank, India’s biggest lender, expects in coming months to launch its own private LLM, which it will use to increase operational efficiency and power customer interface and knowledge management applications, including a virtual assistant for clients.

BloombergGPT, launched in spring 2023, was the first LLM created expressly for the finance industry. Integrated into the Bloomberg Terminal, it performs a range of labor-saving co-pilot functions, like swapping in ticker symbols for company names.

“We’ve just scratched the surface of potential transformations enabled by generative AI, and expect that within the next year, it will gradually integrate into the operations and products of financial institutions and merchants globally.”

Ken Moore,

Mastercard’s Chief Innovation Officer

Banks are focusing on internal gen AI applications — software development co-pilots, knowledge bots, operational efficiency drivers — that serve as testbeds, laying the groundwork for what's to come.

Immediate focus - Experimentation

The next phase will likely involve constructing the architecture for more ambitious gen AI initiatives, such as customer onboarding solutions — while remaining within a proof-of-concept context.

Short-term - Building foundations

Anticipated efforts include producing applications that redefine customer interactions, like client-facing AI financial advisors. This could be contingent on better regulation: Developers will want to know the rules of the game, given gen AI's risks (see "Regulatory guardrails for safe passage" below).

Mid- to long-term - Scaling up

Zero data retention rules

These ensure that LLMs retain information submitted to them in queries only for as long as it’s needed to generate an answer; after that, it’s deleted.

“Generative AI will only become truly useful if we can trust it, and the only way that can happen is to ensure we apply basic principles of responsibility to it and the data that is used to create it.”

Andrew Reiskind,

Mastercard’s Chief Data Officer

Prompt engineering protocols

Prompts to LLMs can be constructed in such a way that they don’t feed personally identifiable and other sensitive information into LLMs.

Data tokenization

The process by which “tokens” bits of data with no inherent value, stand in for sensitive information — a near-time solution to data leakage.

Layered defenses

Bad actors are perpetually devising new ways to “jailbreak” LLMs to produce harmful content or leave malware. Layered defenses incorporating several of the methods discussed here could become critical to combatting such attacks.

Data encryption methods

Banks must employ strong encryption methods for data at rest and in transit. Though currently in its early stages of development, homomorphic encryption, a technique that allows computation to be performed on ciphertext, could play a role here in the future.

Cloud-hosted modifications of proprietary LLMs

Banks are acquiring licenses for LLMs and setting up AI systems based on them in public clouds; in those secure environments, with appropriate security controls, the LLMs are fine-tuned on bank data to fit the specific banks' needs.

Protective data firewalls

These are becoming standard, ensuring that AI systems stay at arm's length from sensitive information yet can still use less critical bank data.

In addition, gen AI’s proficiency in processing unstructured as well as structured data means that it could give bankers faster, more frictionless access to information sources that previously took significant human effort to mine for insights. These sources might be internal ones like quarterly reports, emails, board meeting transcripts and consumer sentiment research, and external ones like governmental transcripts, think tank white papers, and economic and political news reports. Again, bankers would receive information from these sources in a clear, digestible format.

Gen AI could also boost the productivity of the systems integrators and other service providers to which banks outsource tasks. As it does, banks could want to renegotiate contracts, expecting faster deliverables at lower prices from these providers. Finally, gen AI’s natural language capabilities, translating IT work into “human” language and vice versa, could make what happens in IT more comprehensible to non-tech bank executives. As these decision-makers gain a more transparent view of IT, the latter may achieve a new strategic importance.

There is also the potential for gen AI's pattern recognition capabilities to be used in detecting irregularities that indicate fraudulent transactions and other threats such as malware — possibly improving the surveillance capabilities of older forms of AI.

Generative AI may enhance customer interactions, promising bots capable of comprehending and responding to inquiries in more contextually appropriate ways. The image of the bank client desperately trying to bypass an automated phone or chat system to reach a human operator could eventually become obsolete. This smooths the transition to a digital-first banking approach, addressing customer service friction points associated with this shift.

The role gen AI could play in augmenting the educational aspects of financial advisory services is notable. Financial advisors and their clients could use AI-powered simulations to deepen their grasp of complex investment strategies like options trading, fostering more informed decision-making. Compellingly, gen AI could democratize high-quality financial advice, extending personalized advisory services to a broader range of people and leveling the economic playing field.

Fueling the journey — talent

Regulatory guardrails for safe passage

Navigating a potential plateau

Today's innovations, tomorrow's foundations

Integrity: Ensuring fidelity to facts

Proportion of Global 2000 companies that will use AI-powered tools to manage their workforces in 2024.

Today, AI-powered automation is already a critical tool for banks' human resource departments. Gen AI may improve their capacities even more. Its ability to handle unstructured data could enrich applicant pools, putting in front of HR managers likely job candidates who may not have posted resumes to employment platforms and who may lack traditional banking employment backgrounds — but who have much to offer nonetheless. Such AI-assisted broadening of the recruitment net could be as positive socially as it could be for banks and the candidates themselves. Gen AI might also create concise summaries of work records and skill sets, enabling human counterparts to gauge the suitability of applicants quickly. Once a new hire joins, generative AI could support onboarding by serving up organizational knowledge upon request.

Talent management

Source: The rise of digital bosses

Please look out for the Q1 2024 issue of Mastercard's thought leadership publication Signals, which will explore six key technological developments to watch.

Analyze and synthesize vast amounts of financial data, enabling banks to gain valuable insights, make informed decisions and offer data-driven solutions.

Automate content creation for marketing, customer communication, and reporting, producing engaging and personalized content at scale.

Improve customer interactions by providing quick and accurate responses, assisting with inquiries and enhancing overall CX.

As for day-to-day IT operations, engineers might tap into gen AI for step-by-step guidance in various tasks, from troubleshooting software to overcoming integration challenges to “explaining” outdated code. Gen AI could function as a pervasive, always-on efficiency driver in automating routine daily tasks and accelerating programming, freeing IT professionals for more critical tasks.

Data privacy: Chinese regulations delineate a unique set of data governance rules influenced by national policies and security concerns.

Andrew Reiskind,

Mastercard’s Chief Data Officer

40, 41

Signal: Chinese fintech Ant Group has created its own LLM. It will power, among other products, an assistant for financial industry professionals and an application that answers customer queries.

Andrew Reiskind,

Mastercard’s Chief Data Officer

The AI-powered bank

Mitigating challenges

Regulatory uncertainties

From exploration to deployment

Knowledge and insights