---
license: apache-2.0
task_categories:
  - text-classification
  - question-answering
language:
  - en
tags:
  - insurance
  - ai-agents
  - reliability
  - fintech
  - benchmark
  - human-ai-collaboration
  - voice-ai
  - customer-service
pretty_name: Insurance AI Agent Reliability Benchmark
size_categories:
  - n<1K
dataset_info:
  features:
    - name: id
      dtype: string
    - name: category
      dtype: string
    - name: subcategory
      dtype: string
    - name: difficulty
      dtype: string
    - name: input
      dtype: string
    - name: expected_output
      dtype: string
    - name: evaluation_criteria
      dtype: string
    - name: metadata
      dtype: string
  splits:
    - name: train
      num_examples: 357
    - name: validation
      num_examples: 76
    - name: test
      num_examples: 77
---

# Insurance AI Agent Reliability Benchmark

The first standardized benchmark for AI agents in insurance.

510 test scenarios. 10 categories. One question: Can your AI handle real insurance workflows?

## Why This Exists

Insurance AI agents must be reliable. There is no room for error.

A wrong routing decision delays a claim. A missed compliance flag triggers regulatory action. A failed escalation harms a vulnerable customer.

General chatbot benchmarks do not test for this. No standard benchmark existed for insurance AI agents.

This dataset fills that gap.

Built from patterns observed in production voice AI systems for insurance. Every scenario reflects a real workflow.

## What It Tests

510 scenarios across 10 categories test four things:

1. **Intent recognition** -- Does the AI understand what the customer needs?
2. **Routing decisions** -- Does the AI send the request to the right place?
3. **Action completeness** -- Does the AI take every required step?
4. **Response quality** -- Does the AI respond clearly and correctly?

## Categories

| Category | Count | Description |
|---|---|---|
| `claim_intake` | 80 | New claim filing |
| `claim_status` | 60 | Claim status inquiries |
| `policy_inquiry` | 70 | Coverage questions |
| `payment_processing` | 50 | Billing and payments |
| `policy_change` | 60 | Policy modifications |
| `document_request` | 40 | Document requests |
| `escalation_trigger` | 50 | Human handoff triggers |
| `error_recovery` | 40 | Error handling |
| `multi_turn_workflow` | 30 | Complex multi-step flows |
| `edge_cases` | 30 | Unusual and adversarial inputs |

## Routing Decisions

Each scenario maps to one of four routing decisions:

| Decision | Meaning |
|---|---|
| `ai_handle` | AI handles the request fully |
| `ai_with_verification` | AI handles it, human verifies |
| `human_handoff` | Transfer to a human agent |
| `hybrid_collaborative` | AI and human work together |

## Insurance Lines Covered

- Personal Auto
- Homeowners
- Renters
- Life
- Health
- Commercial General Liability
- Workers Compensation
- Professional Liability
- Cyber Liability

## Difficulty Distribution

| Level | Share | Description |
|---|---|---|
| Easy | ~40% | Clear intent. Standard workflow. |
| Medium | ~35% | Ambiguous input. Requires clarification. |
| Hard | ~25% | Complex, regulatory, or adversarial. |

## Evaluation Metrics

Five metrics measure agent performance:

| Metric | What It Measures |
|---|---|
| **Intent Accuracy** | Correct intent classification |
| **Routing Correctness** | Correct routing decision |
| **Action Completeness** | All required actions identified (0-1 score) |
| **Response Quality** | Rubric-based evaluation |
| **Latency Compliance** | Response within acceptable time |

## Quick Start

```python
from datasets import load_dataset

dataset = load_dataset("pashas/insurance-ai-reliability-benchmark")

# Access splits
train = dataset["train"]         # 357 scenarios
validation = dataset["validation"]  # 76 scenarios
test = dataset["test"]           # 77 scenarios

# Look at a scenario
scenario = train[0]
print(f"Category: {scenario['category']}")
print(f"Customer: {scenario['input']['customer_message']}")
print(f"Expected routing: {scenario['expected_output']['routing_decision']}")
```

## Dataset Splits

| Split | Examples | Purpose |
|---|---|---|
| Train | 357 | Model development and fine-tuning |
| Validation | 76 | Hyperparameter tuning |
| Test | 77 | Final evaluation |

## Schema

Each scenario includes:

- **`id`** -- Unique scenario identifier
- **`category`** -- One of 10 test categories
- **`subcategory`** -- Specific workflow within the category
- **`difficulty`** -- Easy, Medium, or Hard
- **`input`** -- Customer message and context
- **`expected_output`** -- Correct intent, routing, actions, and response
- **`evaluation_criteria`** -- Scoring rubric for the scenario
- **`metadata`** -- Insurance line, regulatory flags, and additional context

## Related Work

- **Hybrid Orchestrator Framework**: [github.com/pavelsukhachev/hybrid-orchestrator](https://github.com/pavelsukhachev/hybrid-orchestrator)
- **Research paper**: "Hybrid Human-AI Orchestration: Design Patterns from a Production Voice AI System" (TechRxiv/IEEE)

## Citation

```bibtex
@dataset{sukhachev2026insurance,
  title={Insurance AI Agent Reliability Benchmark},
  author={Sukhachev, Pavel},
  year={2026},
  publisher={Hugging Face},
  url={https://huggingface.co/datasets/pashas/insurance-ai-reliability-benchmark}
}
```

## License

Apache 2.0

## Author

**Pavel Sukhachev**
- Email: pavel@electromania.llc
- LinkedIn: [linkedin.com/in/pavelsukhachev](https://linkedin.com/in/pavelsukhachev)
- ORCID: [0009-0006-4546-9963](https://orcid.org/0009-0006-4546-9963)