Business Bench
Business Bench tests AI agents on everyday business work: reconciling spreadsheets, preparing files for import, writing reports and handling routine finance tasks. Each task comes with its input files and a set of automatic checks, and an answer counts only if it passes all of them.
Latest results
We gave three AI agent setups all 187 tasks, three times each, and checked every answer with the same rules. Proto is the agent built by ERP.AI; Codex is OpenAI’s coding agent. Two of the setups use the same underlying model, GPT-6 sol, so the difference between them comes from the agent itself.
- Proto and Codex on the same model are close. Proto got 1.4 points more answers accepted, but the plausible range runs from −1.8 to +4.8, so neither is clearly better on these tasks.
- Proto on DeepSeek V4.1 Flash did as well as either. It also took the least time per task.
- Getting it right every time is harder. Each setup solved roughly four in five tasks on all three tries; the rest were solved only some of the time.
What is tested
| Kind of task | What the agent does | Results so far |
|---|---|---|
| File tasks 187 tasks | Gets a folder of business files and a short request, and produces the finished files: a reconciled workbook, an import file, a summary report. | Latest results and earlier results |
| App building 20 tasks | Builds a small internal business app, then makes three changes to it. | Tasks published; no scores yet. |
| Working inside an ERP 9 tasks | Takes a job such as a buyer or accounts clerk inside a business system for several simulated days. | Early pilot on 6 tasks. |
How answers are checked
A report can look finished and still be wrong: a balanced workbook may leave out a customer, and a correct payment may break an approval rule. So each task lists exactly what a correct answer must contain, and a program checks it: the right rows, the right numbers, formulas that recalculate, and the facts a written summary has to mention. An answer is accepted only if every check passes. The tasks and the checks are public, so anyone can see what was asked and dispute a check.
Earlier results
On 16 September we compared Proto on DeepSeek V4.1 Flash with Codex on GPT-5.6 sol over the same tasks, three times each. Proto got 507 of 561 answers accepted and Codex 473. Those two setups used different models, so that comparison mixes the agent and the model. See the earlier results.
Keep in mind
- ERP.AI builds Proto and runs this benchmark. The tasks, checks and every recorded answer are public so others can check the results; the audit page lists what can and cannot be verified.
- The tasks were available while Proto was being developed, so these results say how the agents do on these tasks, not on work they have never seen.
- Passing most checks is common; passing all of them is harder. The findings page shows how close the failed answers came.
Check it yourself
| If you want to know | Start here |
|---|---|
| What an agent was asked to do | An example task |
| Whether the numbers add up | Verify the results without running any agent |
| How it was measured | Methods and the paper |
| How to run it on your own agent | Instructions on GitHub |
Cite
@misc{businessbench2026,
title = {Business Bench: Evaluating Agents on Business Deliverables},
author = {Misra, Somesh and Misra, Somnath and Dixit, Shashank},
year = {2026},
howpublished = {\url{https://businessbench.org}},
note = {Version 1.0.0}
}Found a problem with a task or a check? Report it on GitHub with the task name.