Should Banks Build their Own LLMs?
Finance giants are pondering if they should challenge Big Tech on the foundational level of AI.
On the go? No problem. We have you covered with the audio version:
Banks around the world are equally exuberant and cautious about generative AI. Financial institutions were among the first to harness the potential of artificial Intelligence. Software was created in-house not only to mimic human behavior but to do tasks like risk scoring or fraud pattern recognition at a level far superior to what humans could do. This was decades ago. Today, the world is enthralled by Gen AI, a subset of deep learning, itself is a highly sophisticated subset of AI. Yet banks are no longer in the lead.
Some banks have merely started tinkering with Gen AI, while others have already embarked on large-scale transformation projects. But none of them has built a large language model (LLM) by itself. LLMs are the heart of Gen AI. They are models that enable Gen AI to understand and create human-like text, speech or code and even enable research into new pharmaceuticals or the human genome. The LLM space is dominated by a handful of companies: OpenAI and Microsoft (ChatGPT), Google (Gemini), Anthropic (Claude), Meta (Llama). Currently, banks are showing very little appetite to join this roster. But that might be changing.
Banks’ different approaches
For individuals, using Gen AI is as simple as opening a webpage or downloading an app. With some applications it even suffices to text a WhatsApp number to receive a GenAI response. For companies, on the other hand, it is a jumbo project. Smaller companies can use the enterprise version of the available LLMs but this is no serious option for licensed financial institutions. Instead, the bare minimum banks have to do is to deploy a self-contained instance of the model on a private cloud.
But this doesn’t quite cut it. If any company wants to gain real value from Gen AI it must embed the LLMs into its existing systems such as Salesforce. Even Gen AI applications such as a customer service chatbot or a coding assistant require this level of integration, so that they can autonomously learn and improve.
For financial institutions, however, the largest value of Gen AI can only be unlocked if the third-party models are fine-tuned. Banks must continue to train those off-the-shelf LLMs with their own data. ChatGPT or Gemini will not be familiar with regulatory reporting requirements or payment flow anomalies unless they learn directly from the bank’s data.
And indeed, most leading banks are opting for this third way, namely to train one of the big LLMs with their own data. The rationale is very simple: Big Tech has spent so much money to get where they are. Even if banks invest billions of dollars, it will take them years to get there and this make them lose the race to those competitors who just build on top of OpenAI’s or Google’s models.
In January 2025 a hitherto unknown Chinese company changed this unquestioned truth. In what has been dubbed AI’s Sputnik moment, the startup DeepSeek released an AI model with a performance close to ChatGPT’s, yet while only using a fraction of the computational power. This bombshell showed many things, but for banks two lessons stood out: First, it is possible to replicate OpenAI’s LLM. Second, you can significantly improve it by building it from scratch. Ergo, a number of American banks is now reported to ponder DIY models replacing third-part software. So, are they on the right track?
More than just another outsourced software program
Most banks think about themselves as tech companies with a banking license. And they are. IT’s importance has grown from a tool for delivering products to a core competence of any financial institution. A key differentiator in the market. A bank’s technology determines profitability, customer acquisition, and loyalty more than anything else. Yet banks outsource their IT with ever more conviction and fervor. Even third-party cloud infrastructure is finally gaining traction. If banks are willing to move their most critical data to Amazon or Microsoft, why bother about outsourcing AI algorithms?
The list of answers is long. The absence of absolute data security, the licensing fees and usage costs, vendor lock-in, a foregone opportunity to license one’s own model to other banks. The architecture of the LLM might not be ideal for the industry use cases or the data pipelines. Furthermore, intellectual property issues with the provider’s training data could leave the bank without an LLM. The operational risk is higher than when licensing, say, a CRM tool.
And in case these aren’t enough reasons, here is the straw that should break the camel’s back: Building your own LLM gives you complete control over the model and its evolution. In the current setup, not only don’t you get to decide how Google or Meta build their algorithms, in most cases you actually don’t even know how they work. Keep in mind that Gen AI, unlike other AI subsets such as decision trees, is not deterministic but probabilistic. It means that you can never be certain about its answer and thus never be certain about the veracity of the output.
Especially with an industry so specific and highly regulated as banking, control is essential. And after all, banks are the real experts in assessing credit risk, making investment decision, or following legal requirements. And they are the ones sitting on all the relevant training data. Is it really that far-fetched to suggest they should be in charge of the foundational models?
There is also an alternative way in which banks don’t have to build an LLM from scratch, but still eliminate most of the drawbacks of a third-party solution. Enter JPMorgan Chase’s LLM Suite, an AI assistant to be rolled out to 140,000 employees. It features document summarization, writing assistance, idea generation, and it acts as a virtual research analyst, but the clever thing is that LLM Suite is not based on a single model. Rather, it acts as a portal that connects to the best-in class models. It still doesn’t give the bank full control over the algorithm, but it significantly slashes all sorts of risks.
Banks’ relationship to Big Tech
How banks approach their AI strategy will also have a far-reaching effect on the industry’s relationship to tech giants more broadly. Today banks accept their reliance on tech giants through gritted teeth. Whether that be Apple or Google Pay, or the oligopoly in cloud computing. They are also intertwined on other levels. A bank can hardly be functional without Windows Office programs, iPhones, or without advertising on Facebook and Google. On the other side, banks are among the largest customers of those tech titans.
Those two groups are mega partners, doomed for collaboration. What is straining this relationship is Big Tech’s foray into the financial services world. Though not obtaining banking licenses directly, they are cutting deep into banks’ margins. How this battle eventually plays out, will be determined by who manages to win and keep the upper hand with transformational tech. AI, digital assets, distributed ledgers. There is no shortage of battlefields. But the strategies are actually quite similar across the tech spectrum: The deeper the layer you are controlling, the better. Check out my book Big Tech in Finance: How to Prevail in the Age of Blockchain, Digital Currencies and Web3 to understand how new tech will reshuffle the financial work and what companies should do to come out on top.
Not subscribed yet?
If you enjoyed this article, why not subscribe for free to the monthly newsletter or invite others to join? Both pay off.
Subscribers to The New Frontier have free access to my 3.5h-long on-demand program that usually costs $99. So if you haven’t viewed the Blockchain and Cryptocurrency Fundamentals seminar yet, give your knowledge a boost now. In case the secret link you received in your registration e-mail has expired, simply take a couple of seconds to re-subscribe. It’s worth the effort!
You can also earn free access to it by referring friends to this newsletter. When you use the referral link below, or the “Share” button on any post, you'll get credit for any new subscribers. Simply send the link in a text, email, or share it on social media.
When friends use your referral link to subscribe, you’ll receive special benefits.
· Get the Blockchain Masterclass for Banking and Finance for free (regular price $99) for 5 referrals
· Get the Blockchain Masterclass for Businesses and Corporations for free (regular price $99) for 10 referrals
· Get a 30 minutes Zoom chat for 25 referrals
To learn more, check out Substack’s FAQ.
Note: None of the content of this article and this newsletter, nor my books, presentations, and seminars is legal, tax, financial, or investment advice. Nothing contained constitutes a solicitation, recommendation, endorsement, or offer to buy or sell any securities or other financial instruments. The content is gathered from publicly available information and the expertise of the author. It does not contain any insider information. Also, as an Amazon Associate I occasionally earn from qualifying purchases.