Startup raises $13M to test AI agents before they work for you

A startup called Plurai has raised $13 million in seed funding to solve a simple problem: AI agents often look great in a demo and then stumble on real work. Its plan is to let them practice on a digital copy of your company's systems before they touch anything real. The company announced the round on Wednesday.

Why agents still need watching

AI agents are programs that carry out multi-step tasks for you, such as booking travel or answering customer questions. They are popular, but the numbers are mixed. LiveBench 2.0 shows Claude Fable 5.2 succeeding on 82.9 out of 100 tasks and ChatGPT-6 Astra on 75.6. In a simulated corporate setting built by the AI infrastructure company EON, the best model tested, Claude Opus 4.8, got tasks right 76.8% of the time.

That means close supervision is still needed. The article gives a few examples of what goes wrong:

• Travel agents can ignore spending policies when customers push back.
• Combining information from several data sets tends to gum up the works.
• A long chat with a customer can overload an agent and send it down wrong turns.

How Plurai's approach works

Plurai builds a digital twin of a company's electronic systems, including the knowledge base and tools an agent would use on the job. It then maps every path the agent could take and uses that map to test for tricky edge cases and fix them. The platform works alongside tools such as Claude Code, Codex, Copilot and Cursor.

The idea comes from self-driving cars. Co-founders Ilan Kadar and Elad Levi worked at Nexar, which develops autonomous driving technology. Kadar worked with Waymo to simulate and evaluate its driving software.

The bigger bet: smaller, cheaper models

Kadar says the real promise is using what the simulations reveal to further train small models tailored to each customer. Tests done with Nvidia and Nebius showed Plurai's models beating mid-tier frontier models at lower cost. Kadar claims the cost can drop by one or two orders of magnitude, meaning roughly 10 to 100 times. That claim comes from the company, not an independent review.

Kadar's argument is that even much smarter models from OpenAI and Anthropic will not know your business and your policy. Reah Miyara, a Google DeepMind researcher and angel investor in Plurai, said IT leaders worry that the cost of top models is not sustainable over the long term.

Who is behind the money

The round was led by Team8, with U&I, Velocity and Mercer also taking part.

What to do

If your business is weighing AI agents, plan to supervise them closely at first. Ask any vendor how they test for mistakes, and whether that testing uses your own systems and rules rather than generic examples.