Enterprise AI, Banking Language Models And Open Finance Agents For Banking Growth

Originally published in . Link.

Stepping out of a comfortable corner and looking around, bank executives evidenced that thousands of banks across the globe are experiencing stagnated growth. Recent innovations, specifically GenAI, stimulate a huge wave of excitement and expectations.

According to a 2023 McKinsey report, increased productivity through GenAI implementation in the banking industry could add between $200 billion and $340 billion in value annually. Is GenAI the silver bullet?

Researcher Carlota Perez wrote that techno-economic paradigm changes at scale seem to recur every 60 years, and each wave offers great new wealth-creating potential. It normally happens in a particular region of the world and unevenly spreads to the rest of the globe through a trajectory of boom and bust, which results in wealth redistribution. The real estate and financial markets seem to be impacted first and foremost by this GenAI competitive advantage.

In short, banks need to embrace AI-powered digital transformation strategically toward open finance—where banking will become much more open and transparent, and AI will be effectively governed and democratized to benefit customers and, hence, the community and society. For example, we’ve created our Digital Trinity solution as a new growth engine for the financial sector and in closely related industries such as real estate markets.

This can be done locally from inside each bank but in collaboration with financial regulatory agents and governing authorities. New monetary policies must be in place to tackle the financial instability and market uncertainty systematically. It is strongly recommended that each bank develop an operable roadmap following our STAGE framework considering strategy, risk, cost, control and talent.

In a nutshell, it is a three-stage journey. Each stage can be built up in parallel, but in planning, we found it to be more effective if it incrementally iterated three sets of transformative activities toward AI-powered banking. Each stage is a substantive undertaking.

1. Building up the enterprise AI foundation.

This included new fine-tuning of the current banking business model toward AI-powered products and services.

2. Building up private or bank-specific LLM (or simply banking language model).

The origin is our Banking Labs LLM.

3. Building up AI-powered and robotics-automated open finance agents.

Open finance instead of decentralized finance is our proposed and preferred approach. We foresee the upcoming new global ecosystem for financial institutions, insurance and real estate markets.

The second stage is the most debatable. This is the question of whether banks should simply use the commonly available LLMs with techniques of prompt engineering, retrieval augment generation (RAG) or parameter-efficient fine-tuning (PEFT), or build their own proprietary LLM to support enterprise AI. Another promising alternative is to leverage increasing longer and larger context windows each major LLM offers, which is to use vectors of tokens to pivot the model with domain-specific data.

It is being debated in every corner of the GenAI space. The reasons are simple but have long tail consequences.

LLMs are compounding the effect of the expected potential to somewhat revolutionize natural language processing tasks in finance. This is not coincidental. Financial institutions accumulate high-quality data and directly engage customers impacted by multimodal interactions, on which LLMs thrive.

More advantageous is that financial institutions can translate investment into rapid returns in the forms of better customer experience, lower operational cost and new business growth. BloombergGPT presents a successful story, which uses 700 billion tokens to build its own foundation models—more than half of which are leveraging financial data.

However, developing proprietary models is not without challenges, which include high costs as well as a lack of talent and established processes. This also depends on the size and quality requirements of the targeting models. We recommend “responsible AI” with a “govern-to-empower” approach to address these challenges to financial institutions in embracing AI and LLM.

We advocate for adopting a data-centric approach to enterprise AI and banking LLM, with substantial consideration of the crucial role of data acquisition, cleaning and preprocessing. With the shifting paradigm of banking toward open finance, AI agents or a robotics workforce will play critical roles in complementing the human workforces with an optimal AI governance model and operational guidelines. With AI-powered new products, platforms and processes, banks can quickly profit from new growth opportunities and effectively mitigate risks.

Conclusion
In conclusion, we propose developing banking language models so that financial institutions can enable enterprise AI as a strategic approach and embrace GenAI for business growth. Our approach is based on three principles: democratization, a data-centric focus and standardized processes using a five-layer framework—which includes a data source layer, a data engineering layer, an LLM layer, an open API layer with gated authentication and authorization, and the application layer.