Plurai Raises $13M to Enhance AI Agents' Reliability through Simulation

The startup Plurai aims to train and evaluate AI agents within digital twins of companies to anticipate edge cases. The company claims that its open models outperform mid-range "frontier" alternatives at a lower cost. It has announced a $13 million seed round led by Team8 to deploy this approach derived from autonomous driving.

Promises of Performance and Cost Attributed to Simulation

According to Plurai, tests conducted with Nvidia and Nebius show that its models surpass mid-level frontier models while being less expensive. Ilan Kadar explains that it is possible to break down a task through simulation, train an open-weight model, and then verify that the quality achieved is superior or comparable to that of a frontier model, with a cost reduction of one or two orders of magnitude. He emphasizes the use of collected data to tailor small models to specific customer needs. Reah Miyara, a researcher at Google DeepMind and angel investor in Plurai, notes that in a context where companies are looking to optimize their AI spending, IT managers find that frontier models are not suited to their specific data and that their costs are not sustainable in the long term. Kadar also believes that, even with advancements, the models from OpenAI and Anthropic will not be aware of the policies and context unique to each company. Plurai hopes that the analogy between AI agents in physical and digital environments will allow for the prediction of these gains.

Agents Still Fragile in Real-World Tasks

Evaluations show that AI agents are generally not reliable enough without close supervision, which recalls their probabilistic functioning in unpredictable environments. For example, an agent responsible for bookings may ignore spending rules, the combination of multiple datasets can complicate responses, and a long customer conversation can overwhelm the system and lead to errors. On LiveBench 2.0, Claude Fable 5.2 scores 82.9 out of 100 tasks, while ChatGPT-6 Astra scores 75.6 out of 100. In a simulated enterprise environment test, Claude Opus 4.8 achieves 76.8% accuracy. In robotics, powerful general AI models are still not deemed reliable enough for widespread deployment, while specialized systems are already in use with clients and collecting field data.

A Method Inspired by Vehicle Autonomy

Plurai plans to expose agents to simulations powered by synthetic data, constructed from the company's digital systems to prepare for edge cases. The team sets up a digital twin that includes the knowledge base and tools accessible in production. Possible agent flows are mapped out and then used to evaluate and correct the model before deployment. The platform operates alongside existing tools like Claude Code, Codex, Copilot, and Cursor.

A Team with Backgrounds in Autonomous Driving and Academic Support

Co-founders Ilan Kadar and Elad Levi have worked at Nexar, which specializes in autonomous driving. Kadar has also collaborated with Waymo on the simulation and evaluation of deep learning-based driving software. He emphasizes that it is easy to produce a demonstration but difficult to achieve a reliable and economically viable solution, hence the creation of a simulation layer. Reah Miyara, a researcher at Google DeepMind and angel investor in Plurai, believes that the company's advantage lies in understanding failure modes in production and generating synthetic data to adjust models.

$13 Million Seed Round Led by Team8, with U&I, Velocity, and Mercer

Plurai announced on Wednesday a $13 million seed funding round. The round is led by Team8, with participation from U&I, Velocity, and Mercer. This funding aims to support the deployment of the simulation platform and the training of small models tailored to the needs of businesses.