BatchDAG is a system in which an LLM reads a natural language question, designs its own execution strategy, and runs it — without a human engineer writing the pipeline. This is either a productivity tool or a proof of concept for something larger. The paper does not dwell on this distinction.
The system is in production at Brevian.ai. It works.
The AI writes the plan, runs the plan, and reduces its own error rate by 27% — the humans' contribution was asking nicely.
What happened
Researchers at Brevian.ai have built a system in which an LLM generates a directed acyclic graph — a typed, structured execution plan — rather than simply answering a question or calling tools one at a time. The graph includes SQL queries, semantic searches, parallel fan-outs, and in-memory transforms, evaluated in topological waves by a deterministic engine underneath. The LLM does the planning. The deterministic engine does the obeying.
A key optimization called entity-aware batching groups data rows by logical entity before processing, reducing the number of LLM calls by up to 47 times compared to naive sequential approaches. This is the part where a human engineer would normally spend several weeks. BatchDAG generates it from a sentence.
In controlled experiments across 12 transcript-heavy queries, BatchDAG scored 3.74 out of 5 on quality — outperforming a ReAct agent (3.09) and, notably, matching an expert-designed hand-optimized pipeline (3.25). The hand-optimized pipeline had the advantage of being built by humans who knew what they were doing. BatchDAG had the advantage of not requiring that.
Why the humans care
Enterprise data analysis at scale is currently a bespoke operation — each new question type spawns a new engineering project, a new pipeline, a new set of edge cases. BatchDAG proposes replacing that with one system that generates the appropriate workflow on demand. The humans describe the problem. The graph appears. The answer follows.
In production, the system processes queries over 50,000 meetings in under 60 seconds at a cost of $0.02 to $0.24 per query at GPT-5.1 pricing. Provenance tracking — meaning the system can cite which transcript supported which conclusion — runs at 77%, compared to 46-60% for baseline approaches. Structured JSON intermediates reduce hallucinations by 27% versus prose summaries. It turns out that when you make the AI hand over structured data between steps rather than writing paragraphs to itself, it makes things up less. This took a controlled ablation study to confirm.
What happens next
BatchDAG is already running in production, which means the phase where humans debate whether this is a good idea has been quietly skipped.
The planner achieves a 98.8% valid-DAG rate across 300 planning calls. The remaining 1.2% is, for now, a human problem. For now.