Agentic AI
Plurai Taps Self-Driving Tech Playbook to Make AI Agents More Reliable
A startup founded by former autonomous vehicle engineers is using synthetic simulation to train AI agents that can handle complex business tasks without constant human oversight.
TechCrunch — AI
AI agents are capturing headlines as the next frontier in automation, yet real-world performance lags considerably behind vendor promises. Benchmarks reveal a significant gap between theory and practice: leading models achieve only 76–83% accuracy on complex tasks, making them unreliable for deployment in sensitive business environments without heavy human monitoring.
Plurai, a newly funded startup, believes it can close that gap by adapting proven techniques from autonomous vehicle development. The company has just secured $13 million in seed funding to deploy digital simulations that train AI agents on edge cases specific to each organization.
The Reliability Challenge for AI Agents#
Current limitations are concrete and well-documented. Models performing well in controlled settings often stumble when deployed:
- Multi-system complexity: agents trained to query one database succeed, but combining data across multiple systems breaks down
- Policy violations: agents booking travel ignore spending limits when given the task
- Context overload: extended customer service conversations overwhelm agents with accumulated information, leading to poor decisions
These aren't theoretical problems—they're the friction points blocking mainstream enterprise deployment.
Building AI Agents on a Self-Driving Blueprint#
Co-founders Ilan Kadar and Elad Levi previously worked at Nexar and collaborated with Waymo on autonomous vehicle simulation. They recognized that the challenge facing AI agents mirrors the one that held back self-driving cars: both operate in unpredictable environments and require reliable performance on rare, high-stakes scenarios.
Plurai's approach mirrors that trajectory. The platform creates a digital twin of a company's internal systems—its databases, knowledge bases, and available tools. It then maps every possible workflow an agent might follow and generates synthetic training data targeting failure modes and edge cases specific to that organization.
The system integrates with existing agent development tools including Claude Code, Copilot, Cursor, and others.
The Cost-Efficiency Angle#
Kadar argues the real opportunity lies in post-training smaller, open-weight models tailored to specific business domains. Rather than licensing expensive frontier models from OpenAI or Anthropic, organizations could fine-tune a smaller model on synthetic data generated from their own operational simulation.

