The key facts

  • Plurai raised a $13 million seed round led by Team8 with participation from U&I, Velocity and Mercer
  • The startup builds a digital twin of corporate systems to prepare agents for rare scenarios
  • Tests with Nvidia and Nebius showed that fine-tuned small models outperform mid-tier frontier models while cutting costs by one to two orders of magnitude

The agent reliability problem

AI agents are heavily hyped, but benchmarks show they are not yet reliable enough to operate without constant supervision. According to LiveBench 2.0, Claude Fable 5.2 successfully completes 82.9 tasks out of 100, while ChatGPT-6 Astra scores 75.6. Testing agents in a simulated corporate environment, conducted by infrastructure company EON, showed that the best model, Claude Opus 4.8, handles tasks with 76.8% accuracy.

An approach borrowed from autonomous driving

Startup Plurai proposes applying to agents the methods that helped bring self-driving cars to the roads: simulations based on synthetic data built from a specific company's digital systems. Founders Ilan Kadar and Elad Levy previously worked at Nexar, a developer of autonomous driving technologies; Kadar collaborated with Waymo on simulation and evaluation of end-to-end driving software. 'When we saw the wave of AI agents, we saw the same pattern: it's very easy to make a demo, but getting a reliable and cost-effective product is hard,' Kadar told TechCrunch.

How the platform works

Plurai creates a digital twin of a company's electronic systems — the knowledge bases and tools the agent will access in production — and then builds a graph of all possible agent workflows and uses it to prepare the model for rare cases. The platform is compatible with existing tools such as Claude Code, Codex, Copilot and Cursor. According to Kadar, the main opportunity is fine-tuning small open-weight models for a specific client's needs: tests conducted in collaboration with Nvidia and Nebius showed that such models outperform mid-tier frontier models while reducing costs by one to two orders of magnitude. Rea Miyara, a Google DeepMind researcher who invested in Plurai as an angel, sees the company's advantage in its understanding of production failure modes and its generation of synthetic data for model tuning.

Funding and the founders' bet

On Wednesday the startup announced it had raised $13 million in a seed round led by Team8 with participation from U&I, Velocity and Mercer. The founders draw an analogy between agents in physical and digital environments: powerful general-purpose models give robots impressive capabilities but are not yet reliable enough for mass deployment, whereas robots with specialized models are already working in the field and collecting real-world data. 'No matter how smart OpenAI's and Anthropic's models become — and they will become much smarter — they will not know your business and your rules,' Kadar said.