Plurai raised $13M to make AI agents more reliable. Its own numbers don't agree with each other.
The startup, co-founded by two Waymo-adjacent engineers, says simulation can fix agents the way it fixed self-driving cars. Its press release and an earlier report cite different savings.
Plurai announced $13 million in seed funding on Wednesday, led by Team8, with U&I, Velocity and Mercer also putting money in. The founders’ pitch, as they told TechCrunch, is that AI agents have the same problem self-driving cars had a decade ago: easy to demo, hard to trust on the long tail of situations nobody planned for. Their fix borrows the same tool the driving industry used. Simulate first. Then ship.
Co-founders Ilan Kadar and Elad Levi both have PhDs, Kadar in machine learning and computer vision, Levi in mathematics, and both spent years at Nexar, a company that builds autonomous driving tech. Kadar worked directly with Waymo on simulating and evaluating its self-driving software. “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,” Kadar told TechCrunch.
What Plurai actually builds
Plurai makes what it calls a digital twin of a company’s systems: the knowledge base, the tools, the policies an AI agent would touch in production. It then runs thousands of simulated users, including adversarial ones, against the agent inside that twin, mapping every path the agent could take and flagging where it breaks. A booking agent ignoring a spending policy, a database agent that answers fine on one table but garbles a join across two, a support bot that loses the thread after a long conversation: these are the failure modes Plurai’s own release describes.
The simulation data then trains small, open-weight models tuned to one customer’s specific agent tasks. In Plurai’s framing, those smaller models can replace a frontier LLM at the center of the agent, at a fraction of the cost, while also running alongside it as an always-on check on every action.
Where the numbers disagree
Here’s where I’d slow down before repeating any of this as settled fact. Plurai’s own PR Newswire release, published Wednesday, says the company cut design partners’ failures reaching real users by 82%, lowered running costs by 97%, and that joint testing with Nvidia and Nebius for one enterprise customer beat OpenAI’s and Anthropic’s frontier models on accuracy at 40 times and 44 times lower cost, respectively.
Crypto Briefing, reporting on the same company, cites a different set of figures: “over 43% fewer evaluation failures than GPT-5-mini,” and roughly 8 times lower cost, numbers that read like they come from an earlier benchmark or an earlier round (Crypto Briefing also references an approximately $10 million round from February 2026 that doesn’t appear in Wednesday’s release at all).
| Source | Reliability claim | Cost claim |
|---|---|---|
| Plurai’s Nov. 11 press release | 82% fewer failures reaching users | 97% lower running cost; 40-44x cheaper than frontier models at comparable accuracy |
| Crypto Briefing | 43% fewer failures vs. GPT-5-mini | 8x lower cost |
I asked myself whether these are just two different tests measured two different ways, which is plausible, benchmarks for “agent reliability” aren’t standardized the way, say, a model’s MMLU score is. But neither article explains which customer, which task, or which time window produced its number, so there’s no way for an outside reader to reconcile them.
That gap matters. Until Plurai publishes a methodology a third party could rerun, every one of these percentages is a number the company chose to report about itself, not an independently verified result.
Note on sources: I read Plurai’s own PR Newswire release and TechCrunch’s reporting, which includes on-record quotes from co-founder Ilan Kadar, in full. Crypto Briefing’s figures are cited above exactly as that outlet reported them; I was not able to verify either set of numbers against an independent benchmark.
Who’s backing it, and what they’re not saying
The round includes more than 20 angel investors, among them CyberArk founder Udi Mokady, and Microsoft 365 Copilot’s CTO, Dr. Eyal Carmi, has joined Plurai’s advisory board in an individual capacity. None of that is a technical endorsement of the reliability numbers above. It’s a signal that people close to enterprise software think the agent-reliability problem is real and worth funding, which is a lower bar than “this platform’s benchmarks hold up.”
Plurai says its platform now has a free tier and plugs into tools builders already use, including Claude Code, Codex and GitHub Copilot, plus enterprise agent platforms like Salesforce Agentforce. That’s a real, checkable claim, unlike the percentage figures above, since anyone can go sign up and see what the free tier actually does.
This isn’t the first time an agent’s failures turned up only after it reached real traffic. See 200,000 requests sent to a government website chasing one quiz answer for what that looked like without a simulation layer in the way.
My read
The underlying thesis, that agents need the same brute-force simulation testing self-driving cars got before anyone trusted them on public roads, is a reasonable one, and it’s not a new idea in this space; eval and guardrail startups have been pitching some version of it for two years. What’s new here is Plurai’s bet on also post-training small models from the simulation data, rather than just scoring an agent and handing you a report. If that part works as described, it’s a real cost story for companies running agents at volume.
But I’d treat every specific percentage in this story, 82%, 97%, 40x, 44x, 43%, 8x, as a number to verify once the platform has outside users, not a settled fact about how good Plurai is. A $13 million seed round from well-regarded investors buys a company credibility to go try. It doesn’t independently confirm the benchmarks the company put in its own press release.
The same gap between a confident claim and what actually checks out showed up a day later in a completely different field: see AI agents found a Maya site that archaeologists had already mapped for what happens when nobody checks the literature first.