Are We Mistaking AI Agents for the Whole Transformation
Financial services is an industry built around knowledge, language, and decision-making. If any sector was ready for artificial intelligence, surely it was banking. Companies moved quickly. They announced partnerships with OpenAI. They rolled out ChatGPT Enterprise. They built ChatGPT-like AI assistants and gave employees access to large language models. Two years into this experiment, the industry is now asking a different question: What needs to be in place around AI before we can actually trust it?
The shift from access to governance
In 2023, the focus was access. Financial institutions rushed to experiment with the tech after ChatGPT demonstrated something that felt almost impossible at the time: a machine capable of understanding and generating human-like language. The first wave was about discovery. The excitement was understandable, but it also created unrealistic expectations. Because the assumption was that if AI could understand language, it could eventually understand finance, which isn’t exactly true.
In 2024, the industry’s language started changing. The idea of AI replacing workers became less prominent. The idea of AI assisting workers became more realistic. Morgan Stanley’s OpenAI partnership became one of the most visible examples. The firm integrated GPT-4 into tools designed to help financial advisors search through its research and access information more efficiently. The interesting part was what the model was not asked to do. It was not managing client relationships. It was not making investment decisions. It was not replacing financial judgment. It was just helping advisors spend less time searching and more time advising. The technology became valuable when it supported expertise, not when it attempted to replace it.
Around the same time, banks and fintechs across the industry began launching their own ChatGPT-like assistants, copilots, and enterprise LLM deployments, first for employees, then for customers, to search information, draft content, answer questions, and streamline everyday work. Again, the biggest opportunity was removing friction from everyday work. Finding information, drafting documents, helping employees move faster. In general, the most successful AI implementations were asking LLMs to make the bank work better.
Related: AI Hype Overshadows Deeper Tech Changes
From novelty to accountability
By late 2025 and into 2026, the conversation had shifted again. Institutions began building AI agents designed to execute specific financial tasks like underwriting, compliance, servicing, and other financial workflows on top of proprietary data. As the AI novelty disappeared, the industry then moved from experimentation to accountability. Investors became less interested in hearing that companies had AI initiatives; they wanted evidence around productivity gains and revenue opportunities.
The most successful deployments in banking have historically relied on structured, high-quality data rather than the unstructured text that large language models excel at processing. This creates a distinct challenge for institutions trying to leverage generative tools. They must clean and organize their data before the models can reliably handle it. This reality exposes a fundamental misalignment in how many companies have approached the technology. They focused on the model itself while treating the surrounding infrastructure as an afterthought.
Those points exposed a reality that was easy to overlook during the excitement. The model was never the hardest part. The governance, data quality, and integration processes required to make AI useful in a regulated environment represent the actual bottleneck for widespread adoption. When the industry finally looks past the interface and the marketing materials, it finds that the operational complexity of deploying trustworthy systems is far greater than the engineering challenges of building them. User Blocked from Accessing Website
