Artificial Intelligence: Opportunities and Risks for the Financial Sector

By Leonardo Gambacorta, Head of Innovation and Digital Economy Unit, and Vatsala Shreeti, Economist, Bank for International Settlements (BIS)

The introduction of large language models (LLMs) has brought artificial intelligence (AI) into the spotlight. The availability of significant computing resources and vast troves of data is unlocking the potential of AI across the economy’s sectors. The financial sector, owing to its high share of cognitively demanding tasks, is among those most exposed to AI. The integration of AI in finance is transforming the ways in which markets operate, institutions manage risks and consumers interact with financial services.¹

The enthusiasm around LLMs is new, but the use of AI in the financial sector is not (Table 1). Traditional analytics, such as if-then rules, have long been adopted across several of the financial system’s functions. They have been used for risk assessment, rule-based credit analysis, portfolio optimisation and fraud detection. Since the 2010s, machine learning (ML) models have also made inroads into the financial sector’s applications in a wide range of use cases, such as credit- and insurance-risk analysis, high-frequency trading, anti-money laundering (AML) and combatting the financing of terrorism (CFT) initiatives.

With the advent of the latest frontier of technology, generative artificial intelligence (GenAI), two natural questions arise: What is new about the capabilities that GenAI offers? How will the risks and opportunities faced by the financial sector evolve with expanding GenAI use? GenAI is distinct from its predecessors in three key aspects: automaticity, speed and ubiquity.² Unlike previous generations of AI, GenAI models can operate and make most decisions independently, without human intervention. They are also capable of making decisions in fractions of a second, owing to their capacity to process vast amounts of data much quicker than humans. Additionally, as they make it significantly easier for humans to interact with computers and code, they can be integrated into everyday activities ubiquitously throughout the economy.

GenAI, and particularly its ability to give structure to unstructured data (e.g., videos, podcasts, music, images and photos), can create new opportunities for the financial sector. The combination of increasing computing power and the use of new forms of data can be useful across all functions of the financial sector, including financial intermediation, insurance, payments and asset management. Financial institutions are already using GenAI to enhance credit scoring, back-end processing, robo-advising, customer support and regulatory compliance. As of 2023, more than 84 percent of financial institutions surveyed by the Institute of International Finance (IIF) used AI in their businesses, and 86 percent were planning to expand their inventory of GenAI models.³

At the same time, the metaphorical AI lunch is not free. As the opportunities offered by AI have expanded, so have the challenges. There are several “micro” risks arising from AI use that affect individual financial institutions. Ubiquitous AI use in the financial sector can exacerbate threats to consumer privacy and cybersecurity. Moreover, most AI models have an inherent “black box” nature, and their predictions cannot be easily explained. They may also propagate the biases of the data on which they are trained. Other concerns include the emergence of data silos, model hallucinations and algorithmic coordination. GenAI models, in particular, are also prone to the problem of “garbage in, garbage out”: the quality of the outputs of these models is only as good as the underlying input data.

However, there are also “macro” risks that affect the stability of the financial system as a whole. As AI use continues to gain momentum, we should remain attentive to the systemic risks it can create. Even with its limited capabilities, early AI use caused flash crashes and financial instability. Notable examples include the 1987 US stock market flash crash caused in part by the reliance on rule-based models by insurance companies.

More sophisticated AI, such as machine learning models, have amplified these risks in several ways. First, most AI models rely on similar datasets. Due to economies of scale and scope in data collection, a small number of major players, often large technology firms, dominate the production of the relevant datasets used to train these models. Using the same underlying datasets can increase the risks of uniformity and pro-cyclicality in the models’ predictions. Second, as financial institutions rely only on a handful of third-party model providers, there is also a risk of “model herding”. Similar models and optimisation algorithms can increase market volatility, raise the likelihood of flash crashes and reduce liquidity during periods of stress.⁴ Third, increasing network interconnectedness in finance and the real economy can compound the detrimental effects of AI on financial stability. To add to these challenges, the lack of explainability inherent to AI models may prevent regulators from spotting systemic risk or market manipulation in time.

The widespread use of GenAI models amplifies some of the same risks. The automaticity, speed and ubiquity of GenAI can further intensify herding and uniformity. Take, for example, LLMs used for client-facing services such as robo-advising, an important application of GenAI in the financial sector. If most robo-advisors rely on the same underlying models, their advice can become increasingly homogenised.

A distinct but related aspect is the impact of increasing market concentration and cyber-risks on financial vulnerabilities. As we have noted, AI models rely on vast amounts of data and sophisticated algorithms to function, and a failure in any part of this system could lead to significant systemic consequences. Concentrated third-party dependence on the same AI providers can create systemically important single points of failure. For example, a widespread data breach, a software bug or an attack on the AI foundational models used across multiple institutions could trigger a cascading effect, disrupting global financial markets.

Beyond the vulnerabilities arising from market concentration and third-party dependence, spillovers from AI use in the real economy could also be detrimental to financial stability. In general, there is much uncertainty about the impacts of AI on the real sector, particularly on labour markets and productivity. Recent research suggests that AI can increase productivity, especially in tasks that require high cognitive skills—particularly of workers who are less experienced.⁵ If AI behaves like other general-purpose technologies, it could raise productivity, create new tasks and increase demands for labour.⁶ On the other hand, AI can also replace workers and tasks.

The overall impact of AI on the real sector will depend on the balance between productivity increases, task creation and job displacement. In the optimistic scenario, AI adoption leads to positive productivity shocks and limited labour-market disruptions. In this case, the impacts on financial stability will be limited. In the disruptive scenario, the capabilities of AI advance very rapidly and cause massive labour-market disruptions and redistributions of wealth, leading to widespread defaults and financial instability. The reality will probably be somewhere in the middle, and it is essential to steer technological developments towards the optimistic scenario.

Ultimately, the challenge for regulators, financial institutions and policymakers is to harness AI’s benefits while mitigating its risks. Ensuring transparency, accountability and resilience in AI models will be critical to maintaining financial stability in an increasingly automated world. The potential for AI to both stabilise and destabilise financial systems means that careful monitoring and regulation will be essential as AI applications continue to evolve within the financial sector.

Looking ahead, significant uncertainty surrounds not only the long-term diffusion and impact of AI but also the evolution of the technology itself. Progress is being made through the development of AI agents—models that can act autonomously, possess long-term memory and exhibit extensive planning capabilities. This advancement represents a step towards artificial general intelligence (AGI)—AI systems capable of performing all the cognitive tasks that humans can perform. While AGI has the potential to further revolutionise the financial sector, the broader economy and society at large, predicting if and when its potential will be reached remains challenging. Despite this uncertainty, the overarching question we need to answer collectively is what kind of technological advancements we desire in the future, which skills and tasks we want to automate and which fundamental rights these technologies should respect for wider social benefits.

References

¹Bank for International Settlements (BIS): “Intelligent financial system: how AI is transforming finance,” Iñaki Aldasoro, Leonardo Gambacorta, Anton Korinek, Vatsala Shreeti and Merlin Stein, June 13, 2024, BIS Working Papers, Number 1194.

²Bank for International Settlements (BIS): “Artificial intelligence and the economy: implications for central banks,” June 2024, “Annual Economic Report,” Chapter III.

³Institute of International Finance (IIF): “IIF-EY 2023 Public Survey Report on AI/ML Use in Financial Services,” Daniel Mendez Delgado, Conan French and Jessica Renier, December 14, 2023.

⁴Organisation for Economic Co-operation and Development (OECD): “Artificial Intelligence, Machine Learning and Big Data in Finance: Opportunities, Challenges, and Implications for Policy Makers,” August 11, 2021.

⁵ Bank for International Settlements (BIS): “Survey evidence on gen AI and households: job prospects amid trust concerns,” Iñaki Aldasoro, Olivier Armantier, Sebastian Doerr, Leonardo Gambacorta and Tommaso Oliviero, April 23, 2024, BIS Bulletin, Number 86.