Scaled Cognition raises $100m to fix enterprise AI reliability

Scaled Cognition, a Mountain View-based AI model laboratory, has closed a $100 million Series A led by Khosla Ventures, with strategic participation from contact-centre platform Genesys. The round targets what the company describes as the fundamental architectural flaw holding back enterprise AI: not capability, but reliability. For cross-sector leaders already deploying AI in consequential workflows, credit decisioning, medical triage, insurance claims, telecom billing, the distinction carries real operational weight.

The company's founders bring pedigree that sharpens the pitch. CEO Dan Roth and CTO Dan Klein, a UC Berkeley professor of natural language processing, previously built and sold one of the earliest agentic AI businesses to Microsoft. Their new flagship model, APT (Agentic Pretrained Transformer), is positioned as a smaller, faster and cheaper alternative to frontier large language models, with the added claim of guaranteed policy-adherent outputs and hallucination elimination. Scaled Cognition calls this "Super-Reliable Intelligence." The model is deployable inside a customer's own cloud environment or on-premise infrastructure, cutting the dependency on third-party API providers that many regulated industries find difficult to accept.

A different architectural bet

The architecture argument is the investment thesis. Vinod Khosla said the company "developed a different approach, then combined it with the best of LLMs," contrasting this with the more common tactic of adding a reliability layer on top of an existing frontier model. The company says it drew on techniques analogous to verifiable reinforcement learning, the method that has significantly improved coding-model accuracy, and applied them to conversational AI. The claim that APT "will not give you a wrong answer" is an absolute one; Disrupts notes that no publicly available third-party benchmark has yet validated that assertion, and enterprise buyers should treat it as a target posture rather than a certified specification.

Genesys, which serves more than 8,000 organisations across 100-plus countries via its cloud contact-centre platform, is already integrating APT for agentic virtual-agent capabilities. Over the next twelve months, Scaled Cognition says companies using its models are on track to automate more than one billion customer service interactions, a projection that, if approached, would represent a material shift in how contact-centre labour is deployed globally.

The BPO disruption play

The longer-term strategic framing points well beyond customer service. Scaled Cognition is positioning APT as the infrastructure layer for enterprises seeking to "insource" business process outsourcing (BPO) work currently managed by third parties. The global BPO market is valued at approximately $600 billion, covering customer service, IT support, HR administration and finance operations. The model: replace external managed-service providers with an owned AI workforce that the enterprise controls directly.

This is a significant cross-sector signal. For the outsourcing and professional-services industry, it represents a structural demand risk. For financial services and insurance, it creates a procurement decision: build an in-house AI labour stack or remain dependent on a BPO provider increasingly displaced by the same technology. For sovereign and institutional investors in enterprise software, it raises the question of which layer of the AI stack captures durable value: the foundation model, the reliability wrapper, or the vertical-specific deployment platform.

Capital is clearly moving toward the reliability layer. Scaled Cognition's raise follows a broader pattern of investor disillusionment with raw capability benchmarks and a pivot toward production-readiness metrics. The market is beginning to bifurcate between AI vendors that win proof-of-concept evaluations and those that survive enterprise production at scale. Scaled Cognition is explicitly betting on the second category, and Khosla's backing suggests at least one top-tier venture firm agrees the architectural differentiation is real. Whether the $100 million is sufficient to compete against well-capitalised foundation-model incumbents also moving aggressively on reliability is the open question heading into the second half of 2026.