A founder who keeps returning to the same question: can machines get the facts right?
Most artificial intelligence demos are designed to dazzle. A model writes a sonnet, drafts an email, answers a riddle. Dan Roth has spent a career watching what happens a few seconds later, when someone checks the details and finds the confident answer was wrong. That gap - between sounding right and being right - is the whole reason his company exists.
Roth is the co-founder and chief executive of Scaled Cognition, a Mountain View, California company building AI models it says are reliable enough to act on their own in places where a mistake costs real money or real trust. In June 2026 the company stepped out of relative quiet with a $100 million Series A led by Khosla Ventures, at a valuation reported to be around $750 million. For most founders that would be a debut. For Roth it is closer to a fourth act.
They can create incredible answers, and then you can ask them the same question a second time and get a completely different answer that might not even be correct.
Dan Roth, on today's frontier AI modelsHe has described frontier models as "schizophrenic geniuses" - brilliant one moment, confidently mistaken the next. It is a vivid line, and it captures the thesis behind Scaled Cognition cleanly. If you want AI that can change a customer's flight, adjust an insurance claim, or read back a bank balance, brilliance is not enough. It has to be dependable. Roth's word for the standard is blunt: AI has to be "provably reliable" before anyone should trust it.
From speech to Microsoft and back again
Roth's route into AI did not start with computer science. He studied biology at Trinity College in Hartford, Connecticut, finishing in the mid-1990s. What followed was a long run of building companies at the edge of speech and language technology. He led Voice Signal Technologies, a speech-recognition company later acquired by Nuance Communications for a reported $300 million, and Shaser BioScience, which was acquired by Spectrum Brands. Different industries, similar instinct: find a hard technical problem with a real market, and build the company that solves it.
In 2014 he co-founded Semantic Machines with UC Berkeley professor Dan Klein and Stanford's Percy Liang, aiming at conversational AI years before chatbots became a household phrase. Microsoft acquired the company in 2018, and Roth stayed on as corporate vice president of conversational AI, helping fold the technology into Microsoft's broader language efforts. He and Klein had, in effect, built and sold one of the first agentic AI companies. Then they went and did it again.
Scaled Cognition put the Semantic Machines partnership back together - Roth as CEO, Dan Klein as CTO - this time aimed squarely at reliability.
The model with an apt name
Scaled Cognition's flagship is the Agentic Pretrained Transformer, or APT. The name is a small joke that doubles as a mission statement - a model built to give the apt answer. The company reports that APT tops agentic benchmarks such as tau-bench and ComplexFuncBench, and pitches it as smaller, faster and cheaper than frontier models while being more accurate on the tasks enterprises actually run. Rather than a general chatbot, APT is meant to power AI agents that follow policy and take action: the kind of work that sits behind a call center or a claims desk.
What APT is built to optimize for
Illustrative summary of the company's stated design priorities, not measured benchmark scores.The deployment picture the company describes is unusually concrete for a young startup. Scaled Cognition says APT is already in production with Fortune 500 firms across financial services, healthcare, telecom and insurance, and it has set an ambitious near-term goal: automating more than one billion customer-service interactions within twelve months. Genesys, the customer-experience software giant, both invested in the round and integrated APT into its cloud platform - a vote of confidence that comes with distribution attached.
You could have an interaction that was spectacular, then discover the system was making grievous errors.
With our models, most issues get fully resolved, resulting in hundreds of millions saved in operational costs.
Why reliability, and why now
Roth makes the stakes tangible with examples that need no exaggeration. In one, a single hallucinated digit in a prescription could endanger a patient. In another, a wrong number on a bank statement erodes the trust a company spent decades building. The argument is not that frontier models are bad - Roth is quick to call them amazing - but that "usually right" is a failing grade for the tasks he is targeting. An agent that resolves a customer's problem most of the time, and quietly invents an answer the rest, is worse than no agent at all.
That framing is also a business bet. While much of the industry races to build ever larger, ever more capable general models, Roth is betting the next phase of enterprise AI will be won on dependability - on models boring enough that a bank, a hospital, or an airline can hand them real authority. It is a contrarian stance dressed in plain clothes, and it is the same instinct that has run through his whole career: understand language, but get the facts right.
AI has to be provably reliable to earn a user's trust.
Dan RothThere is something fitting about a repeat founder returning, a decade later, to the partnership that first worked. Roth and Klein already proved they could build conversational AI ahead of the curve and find a buyer in Microsoft. The unfinished business - the part the market wasn't ready for and the technology couldn't guarantee - was trust. With APT, a $100 million balance sheet, and a roster of enterprise customers already leaning on the model, that is the problem Roth has set out to close. Whether Scaled Cognition becomes the reliability layer for enterprise AI is still an open question. But few founders arrive at it with as much scar tissue, or as clear a target, as Dan Roth.