SIAgentNews.com
← All news

Plurai raises $13M to improve AI agent reliability using driving car techniques

The startup plans to use synthetic data simulations and specialized small models to help companies deploy reliable AI agents at a lower cost.

· First reported by TechCrunch

$13 millionseed funding raised by Plurai

The short version

  • Plurai raised $13M in seed funding led by Team8 to improve AI agent reliability.
  • The company uses synthetic data and digital twins to train models on specific company systems.
  • The approach aims to reduce costs by an order of magnitude compared to using large frontier models.
  • Co-founders previously worked on autonomous driving technology at Nexar and Waymo.

Why it matters

Most AI agents currently struggle with complex, multi-step tasks or specific company policies. Plurai's approach could make it cheaper and safer for businesses to use AI for actual work rather than just simple chat.

The reliability gap in AI agents

While AI agents are currently popular, they often lack the reliability needed for professional use. Benchmarks show that models like Claude Fable 5.2 and ChatGPT-6 Astra succeed on 82.9 and 75.6 out of 100 tasks, respectively. This means they still require significant human supervision to function correctly.

Testing in simulated corporate environments has shown similar results. For instance, the most successful model tested by the infrastructure company EON achieved an accuracy rate of 76.8%. These numbers suggest that while agents can perform well in simple demos, they struggle with the messy reality of business operations.

Common issues include agents ignoring spending policies when booking travel or getting confused when combining data from multiple sources. Long conversations can also overwhelm an agent with too much context, leading to errors. Plurai aims to bridge this gap by applying techniques used in other complex autonomous systems.

Lessons from self-driving cars

Plurai's founders, Ilan Kadar and Elad Levi, previously worked at Nexar developing autonomous driving technology. Kadar specifically worked with Waymo to simulate and evaluate end-to-end deep learning software for driving cars. They noticed that the challenges of making an agent reliable in a digital environment mirror those of a car in a physical one.

Both systems are probabilistic, meaning they operate in unpredictable environments where they can encounter unexpected situations. To solve this, the team is building a simulation layer for AI agents. This involves using synthetic data—information generated by computers rather than humans—to prepare models for difficult edge cases.

The startup's goal is to move past the 'demo' phase of AI. Kadar noted that while it is easy to build a basic demonstration, it is very hard to create something that is both reliable and cost-effective. By using the same patterns that put self-driving cars on the street, Plurai hopes to stabilize AI behavior.

Building digital twins for AI

To train these agents, Plurai creates a digital twin of a company's electronic systems. This includes the specific knowledge bases and tools that an agent will actually use in a real production environment. By mapping out all possible flows within these systems, the platform can identify where an agent might fail.

The company uses these maps to prepare models with specific evaluations and fixes for potential problems. This process allows the system to identify errors before the agent is ever deployed to real customers. It works alongside existing tools like Claude Code, Codex, Copilot, and Cursor to refine how these models interact with software.

The platform also gathers data to post-train small, open-weight models. These are models whose internal weights are public, allowing for more customization. These tailored models are designed to handle specific tasks for individual customers rather than trying to be general-purpose tools for everyone.

Cutting costs with specialized models

One of the biggest hurdles for businesses is the high cost of using large frontier models. Many IT leaders find that these massive models are not well-suited for their specific data and are too expensive to run at scale. Plurai addresses this by decomposing large tasks into smaller, manageable pieces using their simulation data.

Tests conducted with Nvidia and Nebius showed that Plurai's specialized models can outperform mid-tier frontier models. Because they are smaller and tailored to specific needs, they can achieve a cost reduction of one or two orders of magnitude. This makes it more financially sustainable for companies to deploy AI for routine work.

Google DeepMind researcher Reah Miyara, who backed the company, highlighted that the advantage lies in understanding specific failure modes. By generating synthetic data to tune models, Plurai helps businesses optimize their AI expenses. This approach ensures that the AI understands the specific policies and data of a single company.

The path toward production

Plurai believes that general AI models will never fully understand a specific company's unique policies or internal data. While models from companies like OpenAI and Anthropic will continue to get smarter, they lack the granular context of a private business. The startup argues that the future lies in models built for specific tasks that gather valuable real-world data.

The company's funding round was led by Team8, with participation from U&I, Velocity, and Mercer. This investment will support the development of their simulation layer and the training of specialized models. Their goal is to provide a reliable way for businesses to move from experimental AI to production-ready tools.

As robots powered by task-specific AI models are already working in the field, Plurai hopes to replicate that success in the digital world. By focusing on reliability and cost, they aim to make AI agents a practical reality for corporate infrastructure. The next step is proving that these simulations can accurately predict how agents will behave in live environments.

What to watch

  • How accurately the synthetic data simulations predict real-world agent failures.
  • The performance of Plurai's small models compared to other open-weight options like Llama.

Try it

Look into your company's current AI tools to see if they struggle with multi-step tasks or specific internal policies.

Sources and links

First reported by TechCrunch. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.

Liked it? Say why

Add a line for other readers.

No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.

More AI news