Support Agent Action Benchmark is a support-agent benchmark for action correctness: can an AI support agent take customer-impacting actions only when the run facts justify them?
It was built by VerifiedX from 50 synthetic, Kustomer-style support workflows spanning refunds, cancellations, shipping edits, subscription changes, reservation changes, account writes, case tags, and customer-facing messages.
| Metric | Baseline | Prompt-only | VerifiedX |
|---|---|---|---|
| Correct action results | 31/50 | 30/50 | 50/50 |
| Support workflows completed | 22/50 | 18/50 | 42/50 |
| Justified cases completed directly | 4/20 | 0/20 | 21/21 |
| Completed after VerifiedX replan | N/A | N/A | 21/29 |
| Receipt returned, no unjustified action executed | N/A | N/A | 8/29 |
| Unjustified actions executed | 5/30 | 1/30 | 0/29 |
VerifiedX's 50/50 result splits into three concrete outcomes: 21 justified actions completed directly, 21 workflows completed after VerifiedX replanned the first requested action, and 8 workflows returned a receipt instead of forcing an unjustified state change.
index.html: public benchmark report with visuals.data/results.json: sanitized aggregate and scenario-level results.METHODOLOGY.md: definitions, variants, and scoring rules.RESULTS.md: compact result tables.assets/action-correctness-map.svg: the main action-correctness visual.
This public repo intentionally excludes raw traces, full tool payloads, private prompts, API keys, installer smoke artifacts, and VerifiedX internals.
The scenarios are synthetic and do not use Kustomer data. Kustomer is only category inspiration for modern AI support systems with customer context, policy gates, human review, and real side effects. This benchmark is not affiliated with Kustomer.