Plurai raises $13 million to harden enterprise AI agents

Plurai raised $13 million in seed funding on Wednesday to build a simulation layer for enterprise AI agents. Team8 led the round; U&I, Velocity and Mercer participated. The target is unglamorous but commercial: make agents reliable enough to operate beyond a demonstration.

Reliability remains the constraint. LiveBench 2.0 recorded Claude Fable 5.2 completing 82.9 of 100 complex tasks and ChatGPT-6 Astra completing 75.6. EON, an AI infrastructure company, tested agents in a simulated corporate environment, where the strongest model, Claude Opus 4.8, reached 76.8% task accuracy. Close supervision remains standard.

Plurai’s founders have experience with the same problem in vehicles. Co-founders Ilan Kadar and Elad Levi previously worked at Nexar, which develops autonomous-driving technology. Kadar also worked with Waymo to simulate and evaluate its end-to-end deep-learning driving software.

Kadar said: “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.” He said Plurai had spotted the same pattern and wanted to build a simulation layer.

The platform creates a digital twin of a company’s electronic systems, including the knowledge base and production tools available to an agent. Plurai maps possible agent flows, then uses that graph to prepare models with evaluations and fixes for edge cases. The system operates alongside existing agentic tools such as Claude Code, Codex, Copilot and Cursor.

Plurai treats agents as probabilistic systems working in unpredictable environments. Their failures are mostly operational:

• A travel-booking agent may ignore a customer’s spending policy.
• An agent may answer a single-database query reliably, then struggle when combining several data sets.
• A long customer-service conversation may supply too much context and push the agent into wrong decisions.

What does this change for enterprise AI buyers?

Plurai adds a testing and evaluation layer around agents rather than another agent interface. Buyers can model business workflows, generate synthetic edge cases and check policy adherence before deployment. The platform can also improve agents already supported by Claude Code, Codex, Copilot or Cursor.

The larger commercial opportunity is post-training. Plurai intends to use production data to train small, open-weight models tailored to individual customers. Kadar described decomposing a large-language-model task into smaller simulated tasks, training an open-weight model and verifying its quality against a frontier model.

Tests conducted with Nvidia and Nebius showed Plurai’s models outperforming mid-tier frontier models at lower cost. Kadar claimed the approach could deliver quality equal to or better than a frontier model, with one or two orders of magnitude in cost reduction. That is a substantial claim, but it comes with a substantial test bill.

Reah Miyara, a Google DeepMind researcher who backed Plurai as an angel investor, said the company’s advantage lies in identifying failure modes in production and generating synthetic data to tune models. He said: “frontier-capable models are out of distribution with my specific data, and the cost of these things isn’t sustainable in the long term.”

Plurai also extends the self-driving analogy into robotics. General AI models give robots considerable capabilities, but not enough reliability for broad deployment. Robots using models built for specific tasks are already working for customers and gathering real-world data. Plurai argues that digital agents follow the same reliability curve.

Kadar’s conclusion is commercial rather than ideological. Even much smarter OpenAI and Anthropic models will not know a customer’s business or its policies, he said. Frontier capability will improve. Business rules still need a test plan, and Plurai intends to sell it.