# BankerToolBench ![Category](https://img.shields.io/badge/Category-Agent-1f6feb) ![Metric](https://img.shields.io/badge/Metric-Weighted%20Rubric-2da44e) ![Tasks](https://img.shields.io/badge/Tasks-100-8957e5) ![Saturation](https://img.shields.io/badge/Saturation-low-d29922) ![Created](https://img.shields.io/badge/Created-2026-555) [![Paper](https://img.shields.io/badge/Paper-arXiv-b31b1b)](https://arxiv.org/abs/2604.11304) [![GitHub](https://img.shields.io/badge/GitHub-Handshake--AI--Research%2Fbankertoolbench-181717?logo=github&logoColor=white)](https://github.com/Handshake-AI-Research/bankertoolbench) [![Dataset](https://img.shields.io/badge/Dataset-HuggingFace-ff9d00?logo=huggingface&logoColor=white)](https://huggingface.co/datasets/handshake-ai-research/bankertoolbench) > **Quick answer:** BankerToolBench tests AI agents on end-to-end junior investment-banking workflows — navigating data rooms and financial-information tools to produce real deliverables (Excel, PowerPoint, Word, PDF). Built by Handshake AI Research with input from 502 investment bankers, across 100 tasks. ## At a Glance **What it tests:** Whether an AI agent can complete realistic junior-banker workflows — pulling information from data rooms and financial tools, then producing a polished, correctly-formatted deliverable. **Why it matters:** Financial services is one of the highest-value, highest-scrutiny domains for agentic deployment; this measures whether an agent can be trusted with real analyst-level deliverables, not just numeric Q&A. **Known limitations:** 100 tasks is a modest sample; grading is rubric-based (weighted criteria) rather than fully automatic, so methodology consistency across evaluators matters. ## What BankerToolBench Measures BankerToolBench was built by Handshake AI Research with direct input from 502 practicing investment bankers, to capture the actual shape of junior-banker work: navigating virtual data rooms, pulling from financial-information tools, and assembling the result into a real office-format deliverable (an Excel model, a PowerPoint deck, a Word memo, or a PDF report). Unlike benchmarks that only check a final numeric answer, BankerToolBench grades the full deliverable against a weighted rubric of expert-defined criteria. A high score indicates an agent can reliably chain multi-step financial tool-use with correct, professional-grade document production — a capability that's directly relevant to real deployment in financial-services back-office and analyst-support workflows. ## Benchmark Specifications | Field | Value | |---|---| | Task category | Agent | | Metric | Weighted rubric score (0–1) | | Number of tasks | 100 | | Saturation | Low | | Created by | Lau, Dücker, Chaudhary et al. (Handshake AI Research) | | Source paper | [Lau et al. 2026](https://arxiv.org/abs/2604.11304) | | GitHub | [Handshake-AI-Research/bankertoolbench](https://github.com/Handshake-AI-Research/bankertoolbench) | | Dataset | [Hugging Face](https://huggingface.co/datasets/handshake-ai-research/bankertoolbench) | ## How BankerToolBench Is Scored Each task is graded against a set of expert-defined rubric criteria, each binary (met or not met) and weighted 1, 3, 5, or 10 depending on importance. A task's score is the weighted fraction of criteria met, from 0 to 1; the benchmark score is typically reported as this fraction averaged across all tasks (shown here as a percentage). ## State-of-the-Art Results | Rank | Model | Score | Source | Date | |---|---|---|---|---| | 1 | [Hy4 Preview](/models/hungyuan/hy4-preview) | 78.6% | [Tencent Hunyuan model card](https://huggingface.co/tencent/Hy4-preview) | 2026-08 | *Scores sourced from published technical reports and model cards. Results depend on harness, prompt format, and effort settings — see each source for methodology.* ## BankerToolBench on Benchgen > No Benchgen results yet — [be the first to run BankerToolBench](/benchmarks/handshake-ai-research/bankertoolbench). ## BankerToolBench vs Other Benchmarks | Benchmark | What it tests | Tasks | Saturation | |---|---|---|---| | **BankerToolBench** | Junior investment-banking agent workflows | 100 | Low | | [JobBench](/benchmarks/li-et-al/jobbench) | Real job-role task completion, general occupations | — | Low | | [OfficeQA Pro](/benchmarks/databricks/officeqa-pro) | Office document QA | — | Low | | [Toolathlon-Verified](/benchmarks/hkust-nlp/toolathlon-verified) | Multi-tool agentic workflows, general domain | — | Low | Use BankerToolBench specifically when evaluating an agent for financial-services deployment; use JobBench or Toolathlon-Verified for broader, domain-general tool-use evaluation. ## Run BankerToolBench on Your Model Benchgen lets teams run BankerToolBench against their own model versions, compare results across runs, and catch regressions in financial-workflow reliability — rather than relying on a single vendor-reported number. ## Related Models & Benchmarks - Top performers: [Hy4 Preview](/models/hungyuan/hy4-preview) - Related benchmarks: [JobBench](/benchmarks/li-et-al/jobbench) · [OfficeQA Pro](/benchmarks/databricks/officeqa-pro) · [Toolathlon-Verified](/benchmarks/hkust-nlp/toolathlon-verified) ## Frequently Asked Questions

What is BankerToolBench?

BankerToolBench is a benchmark testing AI agents on end-to-end junior investment-banking workflows, built by Handshake AI Research with input from 502 investment bankers, across 100 tasks.

What does a good BankerToolBench score look like?

As of August 2026, frontier models score in the 70–82% range; scores above 78% represent strong, deployment-ready reliability on junior-banker workflows.

Who created BankerToolBench?

BankerToolBench was created by Handshake AI Research (Lau, Dücker, Chaudhary et al.), published in 2026.

BankerToolBench — Results

BankerToolBench — Phases

BankerToolBench