NOW as Enterprise AI Industry Gold Standard
An open benchmarking suite for enterprise AI agents.
Live site » · EnterpriseOps-Gym leaderboard · EVA-Bench leaderboard
NOWAI-Bench is a coordinated, multi-benchmark effort by ServiceNow to measure whether AI agents perform reliably across the workflows, modalities, and governance demands of real enterprises. Rather than a single test, it is an expanding portfolio of benchmarks — each targeting a distinct slice of what enterprise agents are asked to do.
The current release covers two slices: EnterpriseOps-Gym evaluates long-horizon task agents across eight enterprise domains, and EVA-Bench evaluates voice agents on both task accuracy and conversational experience. Together they span text-based multi-step workflow execution and governed voice interaction — two of the most common deployment patterns for enterprise agents today.
Grounded in production-grade enterprise tasks, NOWAI-Bench provides a rigorous, open standard that helps enterprises make informed model-selection decisions, validate AI deployments with confidence, and meet emerging requirements for AI transparency and accountability.
Long-horizon task agents evaluated across eight enterprise domains (IT, HR, Finance, Customer Service, Procurement, and more).
Repo · Project page · Paper · Dataset · Leaderboard
Voice agents evaluated on task accuracy (EVA-Accuracy) and conversational experience (EVA-Experience), targeting enterprise contact-center and service-desk scenarios.
Repo · Project page · Paper · Dataset · Leaderboard
Snapshot synced 2026-07-10. See the live leaderboards for the full, always-current rankings.
Task Success Rate · Oracle mode
| Rank | Model | Organization | Score |
|---|---|---|---|
| 1 | Claude Fable 5 (with fallback) | Anthropic | 51.1% |
| 2 | Gemini 3.5 Flash | 50.1% | |
| 3 | GPT-5.5 (xhigh) | OpenAI | 46.6% |
| 4 | Qwen3.7 Max | Alibaba | 45.0% |
| 5 | Claude Sonnet 5 (max) | Anthropic | 44.7% |
| Rank | System | Configuration | Pass@1 | 95% CI |
|---|---|---|---|---|
| 1 | Scribe v2.2 Realtime + GPT-5.4 + Eleven Flash v2 (ElevenAgents) | Mixed Models · Cascade | 0.67 | 0.62–0.72 |
| 2 | Universal 3.5 Pro + GPT-5.4 + Sonic 3.5 | Mixed Models · Cascade | 0.60 | 0.55–0.65 |
| 3 | Grok Voice Think Fast 1.0 | Grok · Speech-to-Speech | 0.59 | 0.54–0.64 |
| Rank | System | Configuration | Pass@1 | 95% CI |
|---|---|---|---|---|
| 1 | Scribe v2.2 Realtime + Claude Haiku 4.5 + Eleven Flash v2 (ElevenAgents) | Mixed Models · Cascade | 0.82 | 0.79–0.84 |
| 2 | Gemini 3.1 Flash Live | Google · Speech-to-Speech | 0.59 | 0.56–0.62 |
| 3 | Grok Voice Think Fast 1.0 | Grok · Speech-to-Speech | 0.57 | 0.53–0.61 |
Each benchmark lives in its own repository with self-contained setup instructions. See the individual repos linked above.
The leaderboard snapshot above is machine-managed. When staging is promoted to
production, refresh it from the synced data and publish it to this repo — see
.claude/skills/sync-nowai-leaderboard/SKILL.md in the
NowAI-Bench Space repo
for the exact steps.