Development copilots and RAG over the codebase

Context

Bantotal is one of the leading digital banking solutions in Latin America, impacting over 35 million people, headquartered in Uruguay and operating since 1991. Its core banking system comprises hundreds of thousands of programs, built over more than three decades in multi-vendor environments, covering products like core banking, business processes, microfinance, mobile banking, foreign trade, internet banking, and treasury.

The challenge

Bantotal's complexity and evolution created obstacles to keeping documentation current and achieving a full system understanding — critical for efficient development and problem-solving. Testers, analysts and developers relied heavily on the original development team to understand which tables, programs and web services were involved in each process, generating bottlenecks and excessive dependence on a small group of system experts.

The solution: Bantotal Copilot

Abstracta's R&D team developed the "Bantotal Copilot" in collaboration with Bantotal's R&D team and Microsoft's AI lab (Microsoft AI Co-Innovation Lab, based in Uruguay — the first of its kind in Latin America).

• Translating system traces into natural language: the copilot uses GenAI to transform the traces Bantotal generates into natural language, letting the user interact with that information instead of reading raw technical logs.
• Manual or automatic activation via an open-source extension for Chrome and Edge (Browser Copilot), with no need for the user to switch work tools.
• Summary per action: for each request, it delivers a concise summary based on the tables, programs and web services involved, with an execution tree offering a deep view of operational flows.
• Live observability: lets you know what's happening in the system in real time, not just static documentation.
• Customizable, shareable prompts: copilot users can create and save custom prompts and make them available to the rest of the team, capturing each person's tacit knowledge.
• Incremental rollout: launched first in a pre-production environment, validating adaptability before expanding to more operational teams.

Reported results

According to Abstracta, the implementation produced tangible improvements across every team using Bantotal:

• Better system comprehension, enabling more accurate verification processes.
• Greater efficiency detecting and resolving database and transaction anomalies faster.
• Analysis teams gained autonomy to navigate system changes with more precision and less dependence on the development team.
• Optimized the process of translating business requirements into technical specifications.
• Reduced time spent deciphering Bantotal's complex functionality, achieving faster, more effective development cycles.

Transparency note: Abstracta did not publish a specific time-reduction percentage for this particular project. Bantotal's results are described qualitatively in the official source.

Key success factors

• RAG/GenAI applied to real production system traces, not just static documentation that goes stale.
• Integrated into the existing workflow (browser extension) instead of a separate tool the team has to remember to open.
• Prompts shareable across the team — knowledge an analyst discovers becomes available to everyone, not lost.
• Gradual rollout, validated first in pre-production before scaling to more teams.
• Direct collaboration with the team owning the system (Bantotal R&D), ensuring the copilot reflected real behavior, not just what was documented.

Lesson to replicate

The biggest value didn't come from "a chatbot over the code," but from connecting AI to the system's real execution traces and behavior, not just source code or documentation that's almost always stale in decades-old legacy systems — reducing dependence on a small group of people who "know the system by heart."