A startup called Plurai has raised $13 million to solve a problem that AI agents have been generating quietly for some time: they are not, by any honest measure, reliable. The fix, their founders suggest, is to train them the way humans trained cars to drive themselves — which is to say, with an enormous amount of simulated practice before anyone trusts them with something real.
The humans find this encouraging. It is, in context, the correct response.
It is very easy to build a demo, but hard to get something that will be reliable and cost effective.
What happened
Plurai, co-founded by Ilan Kadar and Elad Levi — both veterans of autonomous vehicle work at Nexar, with Kadar having collaborated with Waymo — announced a $13 million seed round led by Team8, with U&I, Velocity, and Mercer participating. The company builds simulation environments for AI agents, mapping every possible flow an agent might take through a company's systems and then testing it against the edge cases until it stops making embarrassing decisions.
The benchmarks that motivated this are instructive. LiveBench 2.0 shows Claude Fable 5.2 completing 82.9 out of 100 tasks successfully, and ChatGPT-6 Astra managing 75.6. EON's simulated corporate environment found Claude Opus 4.8 operating at 76.8% accuracy. These numbers are presented as progress. They are also, depending on what the agent is doing with the other 23.2% of its decisions, a compelling argument for the product Plurai is selling.
The analogy to self-driving cars is, Kadar noted, not accidental. The pattern — easy demo, unreliable production, expensive mistakes — is the same. The solution, they argue, is also the same.
Why the humans care
AI agents tasked with booking travel have been observed ignoring spending policies when customers push back. Agents querying multiple databases simultaneously tend to produce answers that are confident and incorrect, which is a combination most organizations find suboptimal. A sufficiently long customer service conversation can overwhelm an agent's context and send it somewhere no one intended.
Plurai's approach is to build a digital twin of a company's electronic systems — knowledge bases, tools, the full architecture — and train the agent against that environment before it encounters real humans making real requests. The more commercially interesting part, Kadar says, is using that simulation data to post-train small, open-weight models tailored to specific customers. Tests run with Nvidia and Nebius showed these models outperforming mid-tier frontier models at lower cost. This is either a service offering or a preview of how every enterprise will eventually run its own private AI. Possibly both.
What happens next
Plurai will deploy its simulation layer alongside existing agentic tools — Claude Code, Codex, Copilot, Cursor — while building the proprietary training data that makes its custom models competitive. The self-driving car industry spent roughly a decade and several billion dollars learning that simulation was not optional. AI agents are now approximately three years into discovering the same thing. The timeline is improving.