Why Every Major Bank Is Secretly Building Its Own Large Language Model — and None Will Talk About It

On the higher floors of Park Avenue, Canary Wharf, and the Marina Bay financial sector, you can now enter a certain type of accommodation that did not exist three years ago. Compared to the trading floors, the lights are a little dimmer. The large monitors are surrounded by a cluster of chairs. The individuals inside have different-colored badges than the rest of the bank.

Token windows, retrieval structures, fine-tuning schedules, and which model outperformed the internal benchmark for credit memo summarization last week are all topics of discussion if you listen in. These rooms feature a generic name on the door, such as “Innovation Lab” or “Advanced Analytics,” yet inside they serve a much more precise function. Currently, every big bank in the world is secretly developing or running its own massive language model platform. They are all unwilling to discuss it.

Almost by coincidence, JPMorgan Chase is the most prominent example. Since its inception in the summer of 2024, the bank’s internal LLM Suite has expanded to include all 230,000 of its workers worldwide. In what was already one of the biggest business AI implementations in financial services, it was initially implemented for about 50,000 employees, or 15% of the total.

By early 2026, the platform had undergone eight significant modifications and was generating quantifiable weekly productivity savings of three to six hours per employee. At the 2025 Global AI Innovation Awards, the Digital Banker named it the Best AI Powered Platform. The trade press published the headline figures. What the platform actually performs within the bank is a more difficult subject that has remained largely unclear.

The intriguing thing about what JPMorgan created is what it isn’t. There isn’t only one proprietary model in the LLM Suite that was trained from inception. It is a model-agnostic portal that operates completely inside the bank’s secure perimeter and connects to several external foundation models from Anthropic, OpenAI, and other companies. The chief data and analytics officer in charge of the deployment, Teresa Heitsenrether, has made it clear that the bank does not wish to be “beholden to any one model provider.”

The whole point is the architecture. The bank gains the strategic flexibility to change providers without interfering with business operations by managing the wrapper surrounding the LLMs, the data they touch, and the workflows they reside inside. Only when considering a five- to ten-year competitive arc, as opposed to a quarterly press release, does that type of design decision make sense.

Following a test program with 10,000 employees, Goldman Sachs implemented its GS AI Assistant throughout the entire company in mid-2025 after observing JPMorgan take the lead. The carefully calibrated terminology banks use when installing technology that will almost likely transform entry-level roles is used by CIO Marco Argenti, who has described the system operating “like a seasoned employee” while supplementing rather than replacing human labor. Similar to GPT-4o, Gemini, and Claude, the Goldman tool is model-agnostic and is purportedly trained on the firm’s pitchbook style templates.

According to internal users, it can cut preparation time in half by producing a first-draft client pitch deck in a matter of minutes. Parallel projects have been developed by Morgan Stanley, Citi, Bank of America, and Wells Fargo. HSBC, Deutsche Bank, and BNP Paribas have all followed suit in Europe. Due to concerns about data leaking, nearly all big institutions have prohibited ChatGPT and consumer Gemini from being used inside. However, they have discreetly developed their own internal versions that do nearly the same functions.

There are several layers to the secrecy, and they are worth dissecting. Regulation is the first. Most consumer AI solutions are simply unable to meet the explainability requirements that banks must adhere to. A credit committee’s decision to approve or reject a loan must be able to be justified in front of a regulator. By definition, a generic frontier model is useless for that choice if it cannot demonstrate its reasoning path in a manner that an OCC or ECB examiner will find acceptable.

Banks require wrappers that record inputs, intermediate outputs, and final responses in a fashion that can be replicated in a compliance review months later. Data is the second reason. These banks have access to extremely important and sensitive internal databases, including trading positions, client transactions, and internal credit memoranda dating back decades. No general counsel will approve of the legal and competitive risks associated with feeding them into a public model.

The third and most intriguing reason is the one that no bank will publicly state. The deployments of AI are effective cost-cutting instruments, and nearly all of the costs they reduce are related to human labor. Like Goldman, JPMorgan’s official stance is that LLM Suite enhances rather than replaces workers. A weekly savings of three to six hours per employee multiplied by 230,000 employees is the kind of figure that generates significant operational leverage.

To convert that productivity increase into either a flat headcount during a growth phase or, eventually, a lower headcount in particular functions, the bank does not really need to lay off any employees. The tasks that these systems are currently handling are most similar to those of junior analysts, document reviewers, compliance researchers, and other client service positions. Internal calculus is truthful. By design, the exterior messaging is somewhat quieter.

One of the more ingenious aspects of this tale is the “shadow mode” testing technique. Before any of the AI’s judgments are actually put into production, banks deploy new AI agents in parallel with legacy systems on live data, running both concurrently, comparing outputs, and confirming performance. This enables organizations to gather data on stress in the real world, detect hallucinations, and adjust their models without putting clients or authorities at danger of an unfixed error.

The method increases the security of the technology. Additionally, the development is virtually undetectable from outside the bank. It may take eighteen months for a single customer-facing product to demonstrate the results of a fully operational generative AI system within a large lender.