Company Profile / Enterprise AI
Scaled Cognition
The company teaching enterprise AI to act - not just talk. Its bet: reliability engineered into the model, not bolted on afterward.
The company teaching enterprise AI to act - not just talk. Its bet: reliability engineered into the model, not bolted on afterward.
Most large language models are trained to predict the next word. Scaled Cognition trained its model to predict the next action. That single shift - from language modeling to action modeling - is the whole argument of a company that raised $100 million in June 2026 to make enterprise AI something a bank, a hospital, or an insurer can actually trust with real work.
Founded in Mountain View, California, Scaled Cognition builds what it calls the first Agentic Pretrained Transformer, or APT. Where a general-purpose chatbot aims to sound helpful, APT-1 is built to do a specific job correctly: process an insurance claim, book a flight, issue a refund - and stay inside the rules while doing it. The company frames the problem plainly. AI has become remarkably capable, but it still makes mistakes and hallucinations that keep it out of the workflows where errors carry real consequences.
The answer, according to the company, is not a bigger model. It is a purpose-built one. APT-1 is described as smaller, faster, and less expensive than frontier models, yet more accurate on the agentic benchmarks that enterprises care about. It ships for on-premises, virtual private cloud, or hosted deployment, and requires no customer-specific fine-tuning to go live. The company calls the result "Super-Reliable Intelligence" - a phrase it treats less as marketing and more as a product specification.
"We built Super-Reliable Intelligence ourselves. That's not a feature. It's a fundamentally different company, and it's why the best AI researchers in the world want to work here."
The company's customers live in the industries where AI has been hardest to trust: financial services, healthcare, telecom, and insurance. These are places where a hallucinated policy detail or an off-script refund is not an embarrassing demo - it is a compliance problem, a lost customer, or a bad outcome. Scaled Cognition says it is already in production with Fortune 500 enterprises across exactly these sectors.
The wedge is customer experience. Through a strategic partnership with contact-center giant Genesys - also an investor - APT is being integrated into Genesys Cloud to power virtual agents, with a stated target of automating more than one billion customer interactions within twelve months. It is a telling detail: the company's biggest distribution partner is also one of the firms writing part of the check.
The technical bet rests on three ideas. First, APT-1 optimizes around action-level objectives rather than token-level ones, refocusing the entire agentic stack on decision-making. Second, because the internet has plenty of text but almost no grounded actions, the company built a fully synthetic training pipeline that pairs conversations with the actions and policies tied to them. Third, it uses agent-to-agent self-play - a reinforcement-learning approach where agents practice against each other in simulation before ever meeting a human.
The Agentic Pretrained Transformer - a frontier model built exclusively for customer experience. Conversational quality comparable to leading LLMs, with hallucinations eliminated and policy adherence guaranteed. Supports text and voice across multiple languages.
A full-stack platform for building, testing, and deploying production-ready agents - including synthetic data generation, action modeling, training tools, and agentic evaluation. No customer-specific fine-tuning required.
APT-1 runs on-premises, in a virtual private cloud, or as a hosted service. For regulated buyers, that means data can stay in-house while the model works out of the box.
The crowded field of enterprise AI agents includes frontier labs whose general models are wrapped into agents, and a wave of agent-platform startups such as Sierra, Decagon, and Cresta. Scaled Cognition's differentiation is that it owns the full stack - the data, the model, the training, and the evaluation - rather than fine-tuning someone else's weights.
That ownership is also a recruiting argument. The company's research and engineering bench draws PhDs from UC Berkeley, Stanford, MIT, CMU, Microsoft, Meta, DeepMind, Amazon, and AI2. In regulated markets, the moat is less about what a model can demo and more about what it can guarantee: deterministic behavior, policy adherence, and data that never leaves the customer's environment.
"APT-1 refocuses the entire agentic stack around decision making and action taking."
The three founders previously built Semantic Machines, a conversational-AI pioneer acquired by Microsoft in 2018. Their earlier work helped power conversational features inside Microsoft products - and, the company suggests, taught them that capability was never the real bottleneck. Trust was.
Former Corporate Vice President of Conversational AI at Microsoft; previously co-founded Semantic Machines.
UC Berkeley professor and one of the most cited researchers in natural language processing.
Co-founder of Semantic Machines; leads finance and operations at Scaled Cognition.
Scaled Cognition sells B2B. It licenses the APT model and its agent-building platform to large enterprises, monetizing production customer-experience automation and strategic partnerships like the Genesys integration. Its deployment flexibility - hosted, VPC, or on-prem - is aimed squarely at buyers for whom data residency and compliance are non-negotiable.
In the broader market, the company sits at the intersection of two waves: the rush to put AI agents into real enterprise workflows, and the growing recognition that reliability, not raw intelligence, is the gating factor. Its contrarian positioning - a specialist model that is smaller than the giants but purpose-built - is a bet that in high-stakes settings, buyers will pay for what a system can promise rather than what it can impress.
The founding team's prior conversational-AI company is acquired, seeding the ideas behind Scaled Cognition.
Dan Roth, Dan Klein, and Damon Pender reunite to build reliable, action-oriented enterprise AI.
The company introduces the Agentic Pretrained Transformer, topping major agentic benchmarks.
Khosla Ventures leads a $100M round with Genesys participating, valuing the company at roughly $750M.
Explore APT-1 and the team behind it through the company's own channels and press coverage.
▶ APT-1 on YouTube ▶ Dan Roth interviews ▶ APT-1 product write-upIt builds enterprise AI agents optimized for reliability. Its flagship model, APT-1, is designed to take correct, policy-adherent actions in customer-experience workflows without hallucinating.
APT is a model architecture trained to predict actions rather than just the next word, using synthetic data and agent-to-agent self-play so it can reliably execute real-world tasks like claims, refunds, and bookings.
CEO Dan Roth, CTO Dan Klein (a UC Berkeley NLP professor), and CFO Damon Pender - the team behind Semantic Machines, which Microsoft acquired in 2018.
A $100 million Series A led by Khosla Ventures in June 2026, with participation from Genesys, at a reported valuation of about $750 million.
Fortune 500 enterprises in financial services, healthcare, telecom, and insurance. APT-1 is available on-premises, in a VPC, or as a hosted service.