Scaled Cognition raises $100M to build reliable AI for enterprises

AI hallucination has a $100M answer — and it comes from two former Microsoft researchers.

Scaled Cognition, an AI startup founded by two former Microsoft researchers, raised $100 million in Series A funding to commercialize a model architecture designed to eliminate hallucinations in enterprise AI applications.

"These frontier models are like schizophrenic geniuses — they can create incredible answers, then you ask the same question and get a completely different answer that might not even be correct," Dan Roth, chief executive of Scaled Cognition, said.

The company's Agentic Pretrained Transformer, or APT, predicts structured objects such as programs and system queries alongside traditional token streams, a departure from large language models that optimize for linguistic plausibility without verifying factual accuracy. The approach works best in narrow enterprise domains where the scope of possible queries is limited, Roth said.

The $750 million valuation and backing from Khosla Ventures, one of Silicon Valley's most prominent AI investors, points to growing demand for reliability-focused AI infrastructure. Genesys, a cloud-based customer experience platform, is already using APT within its Genesys Cloud platform for virtual agent capabilities.

Large language models from OpenAI, Anthropic and Google have demonstrated capabilities across general knowledge tasks, scoring above 90% on benchmarks such as MMLU and HumanEval. But those same models can produce confident-sounding but factually wrong answers — a flaw that becomes unacceptable in regulated industries where a single digit error in a prescription number could trigger liability. "A single error can have disastrous consequences," Roth said, describing an automated healthcare agent that "can't afford to hallucinate so much as a single digit in a prescription number."

Scaled Cognition's architecture addresses this by routing different portions of a query to the most appropriate system, depending on the need for reliability, according to Vinod Khosla, founding partner of Khosla Ventures. "It is a separate model for those parts of the system that need real reliability and cannot be subject to hallucination," he said.

The challenge of verifying AI-generated output at scale is a key barrier to enterprise adoption. While a human can easily check a few lines of code, verifying hundreds of thousands of AI-generated lines is practically impossible, said Ion Stoica, a computer science professor at the University of California, Berkeley and co-founder of Databricks. "This makes programmatic reliability an absolute necessity for enterprise systems," he said.

Scaled Cognition plans to use the funding to expand its research team and accelerate enterprise deployments. The Mountain View, California-based startup was founded by Roth and Dan Klein, a natural language processing researcher and professor of AI at UC Berkeley. The pair previously sold their startup Semantic Machines to Microsoft in 2018. The company is targeting customer experience as its first market, a segment where AI-powered agents handle millions of interactions daily and where accuracy directly affects customer retention and regulatory compliance. Genesys's adoption of APT offers an early proof point for the technology's commercial viability.

For investors, the bet on Scaled Cognition reflects a broader thesis: as enterprises move beyond experimentation with generative AI, the ability to trust model outputs becomes a competitive advantage. Companies that solve the reliability problem could capture a disproportionate share of the enterprise AI market, which Gartner projects will reach $300 billion in spending by 2027. Scaled Cognition's narrow-domain approach may limit its total addressable market compared with general-purpose models, but the premium that enterprises will pay for provably correct outputs could offset that constraint.

This article is for informational purposes only and does not constitute investment advice.