As GenAI technology matures and current generation models become cost-effective, three core use cases have emerged. Agents remains the cutting-edge use case, with many firms likely to adopt fast-follower strategies: a single public-facing deployment of an independent AI agent system (leveraging human-on-the-loop controls) by any financial institution would be a substantial inflection point, marking the true start of the race to deploy them across the industry. The odds of that happening in 2025 are high.
Meanwhile, retrieval augmented generation (RAG) and natural language processing (NLP) will continue to dominate the use case space for GenAI, offering straightforward value with straightforward controls.
1. What’s Stalling AI Agents in Finance?
Independent AI agents have not yet been deployed within the financial services and insurance industry, for two reasons: technical limitations and governance limitations.
The technical limitations are connected to the reality that only models from GPT-4 and beyond are effective enough for agentic workflows. These models are expensive to run on a per token basis, and agentic workflows themselves require large amounts of tokens to execute successfully. Of the firms which have been able to deploy high-throughput, production-ready large language model (LLM) pipelines, they have tended to use cheaper and older GPT-3.5 era models. As GPT-4+ models become more affordable and more common for production use during 2025, this constraint will fall away. Firms already capable of deploying these more recent models can move immediately to attempt AI agent workflows, once they address the other constraint: governance.
The governance limitations are fuzzier. There are plausible ‘red lines’ beyond which firms in the industry cannot cross, most notably unconstrained interaction between a customer and an AI agent (human-out-of-the-loop). However, even before reaching those lines there are unclear controls and training requirements for use of AI agent supported (human-on-the-loop) interactions with customers. The first movers in the industry will bear the burden of uncertainty regarding regulatory and public response to these workflows being put into production. Given the substantial efficiency gains available this constraint will not prevent firms from moving forward but it will narrow the pack to only those willing to be first, not followers.
2. RAG vs. Fine-Tuning: The Smarter Choice
The costs and technical burden posed by fine-tuning models has become well understood over the past year, alongside the ease and efficacy of RAG. With context windows in the millions of tokens, investment in a current-generation off the shelf LLM makes more sense in most situations than investing in fine-tuning. While both techniques will continue, and specialized fine-tuned models will have their place, the default means of turning a generic LLM into a subject matter expert for a particular team’s needs will be RAG. Enhanced RAG techniques are emerging, but for most use cases in 2025, ‘naive’ RAG, a well structured prompt, and a reasonable quality knowledge base are sufficient. Dataiku Answers allow these workflows to be set up and maintained with ease: data connection, knowledge bank creation, chat interface, and scalable deployment within a few clicks.
3. LLMs: The New Standard for NLP
One of the most straightforward, easily achievable uses for current LLMs is in providing enhanced NLP. Workflows which previously used older LLM forms (e.g., BERT) or contextless image-to-text conversion (e.g., OCR) can be substantially enhanced by ‘swapping out’ the existing infrastructure for modernized LLM pipelines. This does require consideration and incremental investment, but with no- and low-code LLM solutions it is straightforward to build parallel processes to evaluate improvement and cutover gently. This work has proven value already in 2024 and will accelerate in 2025, becoming a standard use case of AI in-house and integrated within vendor provided point solutions.
Summing Up: AI Opportunities in 2025
The most common use cases in 2025 will be RAG and NLP-based, with internal chatbots connected to large knowledge banks being the most visible and LLM enchanted data structuring pipelines delivering much value in the background. The most exciting use cases around AI agents will emerge, though only from firms willing and able to operate at the forefront of GenAI technology.