Startup Plurai announced a $13 million seed funding round on Wednesday to help enterprise AI agents operate more reliably in unpredictable corporate environments. The funding was led by Team8, with additional participation from U&I, Velocity, and Mercer. Founded by veterans of autonomous vehicle technology, Plurai aims to solve the reliability issues plaguing artificial intelligence agents by borrowing techniques traditionally used to test self-driving cars on public roads.
Despite widespread enthusiasm, current AI agents often struggle with complex multi-step tasks and lack the reliability required for unsupervised deployment. Benchmarking data highlights these limitations. LiveBench 2.0 indicates that models like Claude Fable 5.2 and ChatGPT-6 Astra succeed on 82.9 and 75.6 out of 100 tasks, respectively. Furthermore, testing conducted by AI infrastructure company EON in a simulated corporate environment showed that Claude Opus 4.8 achieved the highest performance among tested models at an accuracy rate of 76.8%.
Key takeaways
- Plurai raised $13 million in a seed funding round led by Team8.
- Co-founders Ilan Kadar and Elad Levi previously worked for autonomous driving company Nexar.
- The startup creates digital twins of enterprise systems to simulate agent workflows and edge cases.
- Tests with Nvidia and Nebius showed custom models outperforming midtier frontier options at lower costs.
Simulating corporate digital ecosystems
Co-founders Ilan Kadar and Elad Levi previously worked for Nexar, a company developing autonomous driving software. Kadar also collaborated with Waymo to simulate and evaluate deep learning driving systems. Recognizing that AI agents face challenges similar to autonomous vehicles, the founders set out to build a specialized simulation layer. Because AI agents are probabilistic systems, they frequently encounter unpredictable scenarios in production. For instance, travel-booking agents may inadvertently ignore corporate spending policies, while multi-database queries or prolonged customer service conversations can overwhelm models with context and cause critical errors.
To combat these failure modes, Plurai constructs a digital twin of a company's electronic infrastructure, incorporating knowledge bases and tools that the agent will access. The platform maps out every potential workflow and utilizes that graph to prepare the model with evaluations and specific fixes for edge cases. While the software integrates with existing developer and agent tools such as Claude Code, Codex, Copilot, and Cursor, Plurai's core offering focuses on utilizing gathered data to post-train smaller, open-weight models tailored to specific corporate needs.
Reducing costs and future outlook
Cooperating with hardware and cloud providers including Nvidia and Nebius, Plurai demonstrated that its custom-trained open-weight models could outperform midtier frontier models while significantly reducing operational expenses. Reah Miyara, a Google DeepMind researcher and angel investor in Plurai, noted that businesses are increasingly finding frontier-capable models expensive and misaligned with proprietary data. By decomposing complex tasks into smaller processes through simulation, Plurai aims to cut costs by one or two orders of magnitude.
Kadar emphasizes that general-purpose AI models developed by firms like OpenAI and Anthropic, regardless of their intelligence, inherently lack specific context regarding an individual business and its internal policies. By drawing a direct parallel to the robotics industry—where specialized, task-specific robots are successfully operating in the field while general humanoid systems continue development—Plurai believes targeted simulation is the key to enterprise reliability.
Comments