S&P Global’s AI arm Kensho has deployed a multi-agent framework called Grounding that consolidates the financial giant’s sprawling data estate into a single natural language interface. The system, built on LangChain’s LangGraph library, routes queries across specialized data retrieval agents covering equity research, fixed income, macroeconomics, and ESG metrics.

For financial professionals who’ve spent hours navigating fragmented databases and learning specialized query languages, the implications are straightforward: ask a question in plain English, get citation-backed answers from verified S&P Global sources.

How the Architecture Works

The Grounding system functions as a centralized router sitting atop what Kensho calls Data Retrieval Agents (DRAs)—specialized agents owned by different data teams across S&P Global’s business units. When a user submits a query, the router breaks it into DRA-specific sub-queries, dispatches them in parallel, then aggregates responses into a coherent answer.

This separation of concerns matters for enterprise deployment. Data teams maintain ownership of their individual agents while the routing layer handles the orchestration. New agents get immediate access to the full breadth of S&P Global data without rebuilding pipelines from scratch.

Kensho’s engineers Ilya Yudkovich and Nick Roshdieh noted that unlike typical web search applications, S&P Global’s data is highly structured and nuanced—requiring more sophisticated retrieval techniques than standard RAG implementations.

The Custom Protocol

Early internal experimentation revealed a common problem in distributed AI systems: inconsistent communication interfaces between agents. Kensho’s response was developing a custom DRA protocol establishing common data formats for both structured and unstructured data returns.

The protocol has already enabled deployment of multiple specialized products—an equity research assistant for sector performance comparison and an ESG compliance agent for sustainability tracking both run on the same data foundation.

What This Signals for Enterprise AI

Three operational insights emerged from the build. First, comprehensive tracing and metadata requirements proved essential for debugging multi-agent behavior at scale. Second, financial-grade trust requirements demanded multi-stage evaluation—measuring routing accuracy, data quality, and answer completeness at each step. Third, continuous analysis of interaction patterns enabled iterative protocol refinement.

The financial services industry has been cautious about generative AI hallucination risks. Grounding’s approach—every response backed by citations to verified datasets—addresses that concern directly. Whether competitors adopt similar architectures will likely depend on how well Kensho’s system performs under real-world query loads across S&P Global’s customer base.

LangGraph, the underlying framework, is an open-source Python library designed specifically for stateful, multi-agent applications. Its adoption by a major financial data provider signals growing enterprise confidence in agentic AI architectures for mission-critical workflows.

Image source: Shutterstock