Researchers propose BatchDAG, an LLM-planned execution system for scalable enterprise data analysis

Researchers at Brevian.ai and collaborators propose BatchDAG, a system that addresses limitations of LLMs in exhaustive, cross-entity analytical tasks over enterprise-scale datasets. Unlike LLMs that struggle with context overflow, loss of per-entity attribution, and linear latency from sequential tool calls, BatchDAG offloads execution to a deterministic engine that evaluates an LLM-generated directed acyclic graph (DAG) of operations.

The system’s core innovation is entity-aware batching, which groups rows by logical entity before parallel fan-out, reducing LLM calls by up to 47x. This optimization directly targets the inefficiency of sequential tool calls that plague traditional agentic approaches in large-scale data analysis.

In controlled experiments on 12 transcript-heavy analytical queries, BatchDAG achieved an average quality score of 3.74 out of 5, outperforming a ReAct agent (3.09/5, p<0.01) and matching an expert-designed pipeline (3.25/5). Provenance tracking also improved, with BatchDAG retrieving transcript evidence for 77% of answers compared to 46–60% for baselines.

A controlled ablation examined the impact of structured JSON intermediates versus prose summaries, finding a 27% reduction in hallucinations when using structured intermediates (paired t-test, p=0.107, n=12). The authors note the ablation’s modest sample size and non-significant p-value, but argue the trend is meaningful for reducing factual errors in generated workflows.

The planner component demonstrated high reliability, producing valid DAGs in 98.8% of 300 planning calls. In production at Brevian.ai, BatchDAG processes queries over 50,000+ meetings in under 60 seconds, with per-query costs ranging from $0.02 to $0.24 at published GPT-5.1 pricing. The authors emphasize that BatchDAG is not primarily an accuracy improvement over hand-optimized pipelines but a general-purpose orchestration layer that replaces multiple specialized workflows with a single, natural-language-driven system.