Plurai Raises $13 Million to Test AI Agents in Simulated Copies of Company Systems

The company is betting that reliability, not raw capability, is what stops businesses from using AI agents without close supervision. Benchmarks bear that out: on LiveBench 2.0, Claude Fable 5.2 completed 82.9 of 100 tasks and ChatGPT-6 Astra 75.6. In a simulated corporate environment built by AI infrastructure company EON, the best model tested, Claude Opus 4.8, reached 76.8 per cent accuracy. For companies paying for agents, the remaining failures mean extra cost and risk.

Co-founders Ilan Kadar and Elad Levi previously worked at Nexar, which develops autonomous driving technology, and Kadar worked with Waymo on simulating and evaluating its end-to-end driving software. Their approach borrows from that field. Self-driving systems were prepared for rare edge cases by running them through large numbers of synthetic scenarios before putting them on public roads.

"The moment that we saw the wave of AI agents, we saw the same pattern: That it is very easy to build a demo, but [hard] to get something that will be reliable and cost effective," Kadar told TechCrunch.

Plurai builds a digital twin of a customer's software environment, including the knowledge base and tools an agent will use in production. It then maps every path the agent could take and uses that map to run evaluations and fix likely failures. Typical failures include travel-booking agents ignoring spending policies when customers push back, agents that stumble when combining data from several databases, and customer-service bots that lose track during long conversations. The platform works alongside existing agent tools such as Claude Code, Codex, Copilot and Cursor.

Kadar said the bigger opportunity is using the data from those simulations to post-train small open-weight models for each customer's task. He said tests run with Nvidia and Nebius showed Plurai's models beating mid-tier frontier models at lower cost. The method, he said, is to break a task into smaller pieces, train an open-weight model on them, and use the simulation to check that quality matches or beats the frontier model at a cost one or two orders of magnitude lower.

That pitch speaks to companies' growing concern over AI spending. Reah Miyara, a Google DeepMind researcher who invested in Plurai as an angel, told TechCrunch that IT leaders are finding that "frontier-capable models are out of distribution with my specific data, and the cost of these things isn't sustainable in the long term."

Kadar argues that general-purpose models will keep improving but will never fully absorb how an individual company works. "No matter how smart OpenAI and Anthropic model will be — and they will be much smarter — they will not know your business and your policy," he said.

Plurai did not disclose its customers or valuation. Its cost-reduction claims rest on its own tests with partners and have not been independently verified.