Plurai Raises $13M Seed for Simulation That Breaks AI Agents Before Customers Do
Half of enterprises have shipped an agent or LLM feature that passed internal evals and still failed in front of customers, according to the VentureBeat research Plurai cites. Plurai's answer is the self-driving playbook: build a simulated world and crash the car there first. The Israeli startup raised a $13M seed led by Team8, with U&I, Velocity, Mercer and more than 20 angels including CyberArk founder Udi Mokady. Microsoft 365 Copilot CTO Eyal Carmi joined the advisory board. Valuation was not disclosed. The platform is now generally available. Release: https://www.prnewswire.com/news-releases/plurai-raises-13m-to-build-the-simulation-infrastructure-for-trustworthy-self-improving-ai-agents-302936422.html
How it works: Plurai builds a digital twin of an agent and its world from the customer's knowledge base, tools and policies, then runs thousands of realistic and adversarial users at it in a sandbox. It looks for policy violations, hallucinations, wrong actions, broken tool calls, failed tasks and data leaks. Once the agent is live, it post-trains small task-specific models so production doesn't depend on a frontier model judging every interaction.
The claims are Plurai's own: development cycles cut from months to hours for design partners, 82% fewer failures reaching real users, 97% lower running cost. In joint work with NVIDIA and Nebius, models it post-trained for one enterprise customer beat OpenAI and Anthropic frontier models on accuracy at 40x and 44x lower cost. Founders Elad Levi and Ilan Kadar come from Nexar, the driving-data company, which explains the simulation-first approach.
Small round, sharp thesis. This week's theme keeps coming back: the verifier is the product, and cost per task gets decided outside the model. Plurai sells both in one box. The test will be whether a twin built from documents catches the failures that come from messy real customers.
← Back to all articles
How it works: Plurai builds a digital twin of an agent and its world from the customer's knowledge base, tools and policies, then runs thousands of realistic and adversarial users at it in a sandbox. It looks for policy violations, hallucinations, wrong actions, broken tool calls, failed tasks and data leaks. Once the agent is live, it post-trains small task-specific models so production doesn't depend on a frontier model judging every interaction.
The claims are Plurai's own: development cycles cut from months to hours for design partners, 82% fewer failures reaching real users, 97% lower running cost. In joint work with NVIDIA and Nebius, models it post-trained for one enterprise customer beat OpenAI and Anthropic frontier models on accuracy at 40x and 44x lower cost. Founders Elad Levi and Ilan Kadar come from Nexar, the driving-data company, which explains the simulation-first approach.
Small round, sharp thesis. This week's theme keeps coming back: the verifier is the product, and cost per task gets decided outside the model. Plurai sells both in one box. The test will be whether a twin built from documents catches the failures that come from messy real customers.
Comments