Vals Secures $40M Series A to Build Closed Industry Benchmarks for Enterprise AI Models

AI evaluation startup Vals has raised $40 million in a Series A funding round led by Andreessen Horowitz, following previous seed backing from 8VC and Bloomberg Beta. The San Francisco-based company focuses on developing industry-specific, closed evaluation systems to test artificial intelligence models across specialized sectors including finance, law, coding, and high-stakes safety domains such as cybersecurity, biosecurity, and international humanitarian compliance. Alongside the round, Vals reported an 8x year-over-year revenue increase, expansion of its workforce from 8 to 25 employees, and the rollout of dedicated evaluation initiatives for federal agencies. This funding underscores a critical bottleneck in the generative AI deployment lifecycle: traditional academic benchmarks (such as MMLU or HumanEval) have become increasingly obsolete for enterprise due diligence. When foundation model providers train directly on web-scale crawl data, public benchmark datasets inevitably leak into pre-training corpora, artificially inflating benchmark scores without guaranteeing reliable execution on nuanced production workflows. By maintaining proprietary, undisclosed test suites across vertical disciplines, Vals offers enterprises and government bodies an independent verification mechanism before deploying models into production. From a cloud and MLOps standpoint, model evaluation has shifted from an offline data science exercise into a continuous DevOps requirement. Modern production pipelines require automated gating mechanisms that assess model drift, recursive self-improvement behavior, safety guardrails, and compliance against contractual and legal standards. The commercial traction of specialized evaluation startups demonstrates that model validation is evolving into an autonomous infrastructure layer—analogous to static analysis, dynamic testing, and vulnerability scanning in traditional software engineering. For platform and AI engineering teams, this development signals the need to move beyond generic synthetic benchmarks when architecting LLM selection and routing engines. Teams operationalizing foundation models should establish closed, task-tailored validation harnesses and implement automated safety evaluation gates within their continuous deployment pipelines, ensuring that model updates preserve domain accuracy, maintain adherence to regulatory constraints, and resist regression over time.