Plurai Raises $13M Seed and Opens Its AI Agent Simulation Platform

Plurai announced $13 million in seed funding and general availability of its AI-agent simulation platform on November 11, 2026, ending a period in which the product reached only select partners and customers. Team8 led the round. U&I, Velocity and Mercer participated, alongside more than 20 angel investors and industry figures.

The investor list carries technical weight. It includes CyberArk founder Udi Mokady and Reah Miyara, senior director of Google Models at Google Cloud and a former model-evaluation lead at OpenAI. Microsoft 365 Copilot CTO Eyal Carmi joined Plurai's advisory board at the same time.

Why does simulation matter for AI agents?

Plurai targets a structural gap in the agent stack: enterprises deploy autonomous agents (software systems that take actions using tools and policies) without a systematic way to test them before customers do. Team8 managing partner Aviad Harell framed the company's bet in infrastructure terms, saying Plurai addresses what is needed to build trust before agents reach customers and to keep improving them in production. "Simulation will become as fundamental to the agent stack as CI/CD is to software," Harell said.

CI/CD (continuous integration and continuous delivery) is the automated testing-and-release pipeline standard in conventional software. Plurai's argument is that agents need an equivalent layer because their failure modes — hallucinations, incorrect actions, broken tool calls, sensitive-data leaks — emerge only under realistic interaction, not static test suites.

How does the platform work?

Plurai builds a digital twin, a software replica of an agent and its operating environment, from the customer's knowledge base, tools and policies. It then runs the agent against realistic and adversarial simulated users inside a sandbox, hunting for:

• policy violations
• hallucinations
• incorrect actions
• broken tool calls
• failed tasks
• sensitive-data leaks

The product page describes isolated sandboxes in which services, documents and world state can be mocked and reset between runs. Production traces — logs of what actually happened in live use — can inform later scenarios.

When the system finds a failure, it traces it to evidence and exact code locations, generates a fix, and validates that fix against a regression set. That last step checks a specific risk: one repair creating another failure elsewhere.

Simulation data also produces training examples for post-training, the additional training of small language models for narrow tasks. Those specialized models can then run the agent itself or act as continuous evaluators and guardrails. Newly detected production failures feed back into simulation, closing the loop. Developers can start from Claude Code, Codex or GitHub Copilot, and teams can use the platform with agents built on Salesforce Agentforce.

What results does Plurai report?

Plurai says design partners cut agent development cycles from months to hours, reduced failures reaching real users by 82%, and lowered the cost of running reliable agents by 97%. The figures come from a select group across AI-native and highly regulated industries, including Fortune 500 companies and major financial firms, before the platform opened more broadly. These are company-reported numbers, not independently audited benchmarks.

An August 15, 2026 customer case study gives the most detailed look at one deployment. An unnamed U.S. enterprise IT managed-services provider built its first custom model in 20 minutes and more than 100 within two weeks. For the portal-routing decision examined in the study, the custom model resolved 89% of requests, against 69% for the customer's existing Claude Haiku setup without reasoning. Latency fell from 1.0 second to 0.15 seconds. Inference cost dropped from $2.87 to $0.076 per 1,000 calls — a reduction of roughly 97% on that specific workload.

A second case study, published October 21, 2026, describes work with NVIDIA and Nebius for a global enterprise customer. Plurai post-trained NVIDIA Nemotron 3.5 Lightning on a training set of 3,722 conversations labeled by a frontier teacher model (a larger model used to generate training labels for a smaller one). The resulting model scored 88.8 out of 100 in quality on real production chats. Plurai compares that with 78.4 for Claude Haiku-5.5 and 88.6 for GPT-6 Luna.

The claim is significant for cost-sensitive deployments: a specialized small model matching frontier-model quality at a fraction of the inference price. It is also self-graded on one customer's traffic.

Where did the approach come from?

Chief executive Ilan Kadar and chief technology officer Elad Levi founded Plurai after working at Nexar, the AI dashcam and road-data company, where Kadar was vice president of AI and Levi an AI architect. Kadar draws a direct line from autonomous vehicles to agents, citing Waymo's more than 20 billion simulated miles — roughly 100 times its real-world driving, by his estimate. The analogy is the company's core pitch: agents, like cars, cannot be validated on live roads alone.

The synthetic-data machinery behind the pipeline comes from BARRED, a preprint submitted April 28, 2026 by Arnon Mazza and Levi. The method decomposes a domain into dimensions and uses multi-agent debate (multiple AI systems checking each other's labels) to verify those labels. It requires only a task description and a small set of unlabeled examples. Plurai says the paper was accepted at ICML 2026, the International Conference on Machine Learning.

The Nexar pedigree and the ICML acceptance signal that the company's approach rests on published research rather than pure product claims — a distinction that matters as evaluation vendors multiply.

What changes with the free tier?

Two things move at once. The seed round funds expansion, and the new free tier converts Plurai from a design-partnership business into a self-service product any team can try. That exposes the company's central unknown to market test: whether the 82% failure reduction and 97% cost savings reported by a small, hand-picked group of design partners survive contact with a broader user base running the full loop of simulation, automated fixes and specialized models.

The stakes extend past one startup. Enterprises face a documented reliability problem with agents in production, and evaluation-and-simulation tooling is becoming the infrastructure layer under that market. If Plurai's numbers hold at scale, simulation moves from optional discipline to standard practice — the outcome Harell's CI/CD comparison anticipates. If they do not, the gap between partner results and self-service reality will define the company's next funding conversation.