DRBench: A Comprehensive Benchmark for Enterprise Deep Research

Résumé

We introduce DRBench, a benchmark for evaluating AI agents on complex, open-ended enterprise deep research tasks. Unlike existing benchmarks focused on surface-level or web-only queries, DRBench assesses an agent’s ability to resolve deep research tasks, which includes resolving high-level queries such as “What changes should we make to our product roadmap to ensure compliance?’’ These tasks are grounded in specific user personas (e.g., product developer) and company contexts (e.g., the ServiceNow organization), and require agents to search across diverse applications accessing both public and private data, reason over heterogeneous formats like spreadsheets, PDFs, websites, and databases, and synthesize their findings into an insightful report. We contribute (1) a suite of complex deep research tasks inspired by ServiceNow spanning three domains: CRM, Strategic Portfolio Management, and Compliance; (2) a realistic enterprise environment composed of several common applications like chat, cloud file system, emails and more; (3) a task generation methodology to generate quality deep research questions and insights grounded on real world web facts and synthesized local facts; and (4) a lightweight evaluation mechanism to score agents’ reports in terms of whether the dominant insights have been extracted and whether the insights in the reports are well-cited. Our experiments benchmark current deep research agents across different methods, revealing performance gaps and observations that would inspire future enterprise deep research agents.