RUBICON: Agentic AI for Messy Enterprise Data

RUBICON is a systems-level answer to the failure of agentic AI on enterprise data. The paper's thesis is that current agentic systems make a category error: they hand the entire query workflow — interpret the request, pick sources and tools, integrate evidence across them, judge completeness, generate the answer — to a single frontier LLM, with text as the representation throughout. RUBICON argues instead for structure: a constrained query interface over each source, and a table-centric integration layer driven by a query processor (a relational optimizer), not a reasoning loop.

It rests on two observations. First, text-to-SQL fails catastrophically on real enterprise data — ~10% accuracy on the Beaver benchmark, rising to only ~30% even when the correct tables and join terms are handed over — so full SQL must be "dramatically subsetted" into what the authors call "baby SQL." Second, integration across disparate corporate datasets is best done over tables, not by flattening maps, CAD, claims, and notes into text for an LLM to reconcile.

The concrete contribution is AQL (Agentic Query Language): a restricted SQL-like notation (FIND … FROM … WHERE …) where the human specifies sources, fields, and intermediate objects, and only the predicate may remain in natural language. Wrappers translate AQL into each source's native interface (database, search engine, e-mail, web API, or LLM) and normalize results into logical relations. Intermediate results are materialized, inspectable, and joinable — so execution is a visible query object, not a hidden chain of tool calls. A human-in-the-loop can inspect and correct the plan.

On RUBICON-Bench (queries coordinating across five heterogeneous sources, modeled on a Munich Department of Transportation deployment), RUBICON achieves 100% end-to-end accuracy while every agentic baseline — vanilla LLM, single-agent ReAct, multi-agent LangGraph — produces zero correct answers in natural-language mode (multi-agent AQL reaches 28.57%). On SemBench, RUBICON beats LOTUS and Palimpzest by 14.7% accuracy, with 62.64% lower latency and 98.64% lower token cost. The takeaway: greater autonomy does not translate into correct multi-source integration; structure does.