# AutomationBench ![Category](https://img.shields.io/badge/Category-Workflow%20Automation-1f6feb) ![Metric](https://img.shields.io/badge/Metric-%25%20Success-2da44e) ![Saturation](https://img.shields.io/badge/Saturation-Low-d29922) ![Created](https://img.shields.io/badge/Created-2026-555) > **Quick answer:** AutomationBench tests AI agents on real business workflow automation tasks — multi-step processes that span multiple applications or systems, similar to what a robotic process automation (RPA) tool or business ops team would handle. Kimi K3 scores 30.8% as of July 2026, reflecting the difficulty of cross-system business automation. ## At a Glance **What it tests:** An agent's ability to automate multi-step business processes across different applications and systems, completing them reliably end-to-end. **Why it matters:** Business process automation is a major commercial use case for AI agents. AutomationBench's low current scores (30.8% for a frontier model like Kimi K3) highlight how much headroom remains for reliable enterprise automation. **Known limitations:** As an emerging benchmark, exact task composition and business-process templates are not yet independently published outside its citation by Moonshot AI. ## What AutomationBench Measures AutomationBench evaluates an agent's ability to automate realistic business workflows that span multiple systems or applications — for example, extracting data from one tool, transforming it, and populating another. Tasks emphasize reliability and correctness across the full multi-step process rather than any single isolated step, making the benchmark a useful proxy for real enterprise-automation readiness. ## Benchmark Specifications | Field | Value | |---|---| | Task category | Agent / business process automation | | Metric | % task success | | Saturation | Low | | Created by | Not yet independently documented | ## How AutomationBench Is Scored Agents attempt multi-step business automation workflows, with outcomes verified against expected end states across the involved systems, producing an aggregate % task success rate. ## State-of-the-Art Results | Rank | Model | Score | Source | Date | |---|---|---|---|---| | 1 | [Kimi K3](/models/moonshot/kimi-k3) | 30.8% | [Kimi K3 technical report](https://www.kimi.com/blog/kimi-k3) | 2026-07 | *Score sourced from Moonshot AI's Kimi K3 announcement, July 2026.* ## AutomationBench on Benchgen > No Benchgen results yet — [be the first to run AutomationBench](/benchmarks/platform/automationbench). ## AutomationBench vs Other Benchmarks | Benchmark | What it tests | Saturation | |---|---|---| | **AutomationBench** | Business workflow automation | Low | | [SaaS-Bench](/benchmarks/platform/saas-bench) | SaaS product usage workflows | Low | | [JobBench](/benchmarks/platform/jobbench) | Job/workforce-oriented agent tasks | Low | | [Tau3 Banking](/benchmarks/sierra-research/3-banking) | Customer-service agentic tasks | Low | ## Run AutomationBench on Your Model Benchgen lets you run AutomationBench against your own agent, tracking cross-system business automation task success rates before production rollout. ## Frequently Asked Questions

What is AutomationBench?

AutomationBench is a benchmark testing AI agents on real business workflow automation tasks spanning multiple applications and systems.

What does a good score look like on AutomationBench?

Kimi K3 reports 30.8% as of July 2026. Given the cross-system reliability required, this reflects a frontier-level result on a still-difficult, unsaturated benchmark.

Who created AutomationBench?

AutomationBench's originating team is not yet independently documented outside of its citation in Kimi K3's July 2026 technical report.

Rank	Model	Score
1	deepseek-v4-1-flash	54.8
2	nex-n2-5-max	50.2
3	nex-n2-5-pro	44.2
4	gpt-6-astra	41.4
5	deepseek-v4-pro	31.8
6	claude-fable-5-1	31.4
7	kimi-k3	30.8
8	qwen3-8-max	27.3
9	deepseek-v4-flash-vision-exp	25.7
10	k2-horizon-375b-a23b	25.3

AutomationBench — Results

AutomationBench — Evaluations