Skip to content
RevensiAIPublic

About

CorpBench Work: 100 deterministic business workflows for evaluating AI agents across finance, marketing, operations, RevOps, and sales.

Topics

Resources

Security policy

Stars

0 stars

Watchers

0 watching

Forks

Repository files navigation

CorpBench Work

Validate public data Apache-2.0 license Hugging Face dataset

CorpBench Work is a public library of 100 deterministic business workflows for evaluating AI agents across Finance, Marketing, Operations, RevOps, and Sales. It describes the work agents were asked to perform, the simulated systems and tool contracts they could use, and the grading mode applied to each workflow.

Explore the live CorpBench benchmark for the interactive leaderboard, findings, cost and speed comparisons, department results, failure modes, reliability, and methodology.

Load the viewer-ready CorpBench dataset on Hugging Face, with separate workflows and results configurations.

Work 1.3 library

Department Workflows Browse
Finance 20 data/work/1.3/workflows/finance
Marketing 20 data/work/1.3/workflows/marketing
Operations 20 data/work/1.3/workflows/operations
RevOps 20 data/work/1.3/workflows/revops
Sales 20 data/work/1.3/workflows/sales

Start with the versioned manifest.json. It lists every workflow, its public path and version, department totals, the benchmark configuration, and SHA-256 digests. Each workflow definition publishes:

  • its objective, systems, limits, and agent-visible tool contracts;
  • deterministic grader metadata;
  • capabilities, version history, and retirement policy; and
  • fixture identifiers and versions, marked withheld.

Seeded workspaces, expected states, reference traces, evaluator implementation, raw run evidence, provider configuration, and private research material are intentionally not published. This repository is a public workflow library and aggregate-data release, not a standalone benchmark runner.

Published results

The machine-readable aggregate behind the live Work 1.3 page is available at results/work/1.3/report.json. It preserves the generated cohort metadata and contains aggregate scores only; it does not contain prompts, trajectories, final workspaces, or run-level evidence.

Data contract and versioning

The current public schemas are in schemas/. Run the repository validator with Node.js 20 or newer:

npm ci
npm test

Library and result paths are versioned as data/work/<version>/ and results/work/<version>/. The immutable tag for this release is work-v1.3; future Work releases use work-v<version>.

Citation

Citation metadata is provided in CITATION.cff. The preferred short citation is:

RevensiAI. CorpBench Work 1.3. 2026. https://github.com/RevensiAI/CorpBench

Project status

This repository is maintained as a read-only public release surface. RevensiAI is not accepting external contributions at this time. Security concerns can be reported using SECURITY.md.

Copyright 2026 RevensiAI. Released under the Apache License 2.0.

About

CorpBench Work: 100 deterministic business workflows for evaluating AI agents across finance, marketing, operations, RevOps, and sales.

Topics

Resources

Security policy

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages