A bank that builds everything on Kotlin, Java and Spring Boot has rolled out an AI assistant using only JVM tools. The team evaluated three libraries—Spring AI, Koog and LangChain4j—and ultimately settled on Spring AI because it fit in "like a native" component with no friction. No new Python hires were required; support stayed within the existing Java development team.
Technically, the solution splits documents into chunks, stores embeddings in PostgreSQL using pgVector, and retrieves relevant context via ordinary SQL queries. The decision to forgo a dedicated vector database such as Milvus or Qdrant was driven by cost and simplicity: the bank already holds a PostgreSQL license and the configuration took only a few days.
The migration assessment shows that licensing costs appear only if premium features are needed, with modest expenses for expanded storage and training the Java team on Spring AI. Compared with typical Python‑centric stacks—hiring a Python engineer, purchasing separate libraries and operating a vector DB—the approach saves roughly 40 % of the project budget.
ROI is already quantifiable: the time spent searching across disparate sources like Excel files, emails and notes has dropped to 30 % of its previous level, delivering a 70 % reduction in search effort. Employee productivity rose by 15‑20 %, accelerating expense approvals and cutting error rates.
What does this mean for a bank CEO whose technology stack is dominated by the JVM? Retrieval‑augmented generation solutions can trim project expenditures by about 40 % and compress budgeting cycles to one‑third of their former duration. Launch a pilot in a single business unit, stress‑test PostgreSQL under load, and devise a training plan for your Java developers.
Why this matters: Leveraging existing JVM infrastructure eliminates the need for new hires and extra services, delivering immediate cost savings. A focused pilot validates performance before scaling, ensuring the bank can reap productivity gains without disruptive investments.