On July 1, 2026, AI model lab Scaled Cognition announced a $100 million Series A led by Khosla Ventures. The round is not about capability. Every major model lab is shipping impressive benchmarks. This round is about the one problem that has quietly blocked most enterprise AI deployments at scale: reliability.
The company’s thesis is precise. Frontier models are genuinely capable. They are not reliably safe to let take real actions, manage real customer interactions, or operate inside regulated workflows where a wrong answer has immediate, measurable consequences. Scaled Cognition was built to solve that specific problem, and Khosla Ventures wrote a $100M check saying the architecture they built actually does it.
The Problem Every Enterprise AI Team Already Knows
Dan Roth, CEO and co-founder, spent years trying to apply AI to business applications before founding Scaled Cognition. His description of what he found is one of the clearest articulations of enterprise AI’s reliability problem to date:
“You could have an interaction that was spectacular, think the singularity is here, and then look at the data and discover the system was making grievous errors. The problem isn’t resources or effort, it’s architecture.”
That gap between impressive demos and reliable production performance is exactly what stalls most enterprise AI programs. A system that works well 70% of the time is a liability when the other 30% means an incorrect insurance claim, a wrong drug dosage lookup, or a billing dispute handled with fabricated policy details. Enterprises are not cautious about AI because they don’t believe it works. They are cautious because they know it does not work reliably enough.
The industry’s standard response has been to add guardrails: output filters, retrieval constraints, confidence thresholds, human review steps. These mitigations reduce harm, but they do not fix the underlying architecture. They add latency, complexity, and ongoing maintenance overhead. And they leave the core failure mode intact: the model can still hallucinate, and the guardrails sometimes catch it, and sometimes don’t.
Scaled Cognition’s claim is that APT does not need those workarounds because the reliability is structural.
What APT Actually Is
APT stands for Agentic Pretrained Transformer. It is Scaled Cognition’s flagship model, designed from the ground up for enterprise operations rather than general-purpose assistant tasks.
The company does not publish a detailed technical paper with this announcement, but the architecture description is pointed. Dan Klein, CTO and UC Berkeley professor of AI, describes the breakthrough as doing “for conversational AI what verifiable reinforcement learning has done for coding.” That framing is significant: verifiable RL for coding (the approach behind systems like AlphaCode and OpenAI’s o-series reasoning for math) works because the domain has ground-truth signals. You run the code and see if it passes tests. Scaled Cognition claims to have found an analogous training signal for conversational tasks, where correctness is harder to verify but still achievable at the architectural level.
The practical properties of APT, as described by the company:
- Smaller and faster than frontier models, meaning lower inference cost at production scale
- Available for VPC and fully self-hosted deployment, with no data leaving the enterprise perimeter
- Policy-adherent by architecture: the model follows rules it is given without requiring constant prompt engineering to reinforce them
- Eliminates hallucinations on in-domain tasks rather than filtering them after the fact
The key competitive claim is the last one. Genesys, which both uses APT inside its Genesys Cloud platform and invested in this round, is the clearest evidence of production deployment. Genesys serves more than 8,000 organizations in over 100 countries. APT is powering agentic virtual agent capabilities across that installed base today.
Vinod Khosla’s Bet and What It Signals
Khosla Ventures is one of the most consequential AI investors operating right now. Its portfolio includes OpenAI (early), Mistral, and several enterprise AI infrastructure companies. Vinod Khosla’s public statements on AI tend to be directionally correct about where the industry is heading, even when the timing is aggressive.
His endorsement of Scaled Cognition is blunt:
“The way to quickly get into the market is to take a frontier model and put a layer on top. What Scaled Cognition did was develop a different approach, then combine it with the best of LLMs. That took more research and more developmental risk. Most people are too lazy to do that. The result is Super-Reliable Intelligence: a model that will not give you a wrong answer. In any industry where an agent takes a real action, nothing else counts.”
That last sentence is the strategic thesis. In agentic deployments, the action is the product. An AI agent that drafts a report has limited blast radius when it’s wrong. An AI agent that adjusts a customer’s account tier, submits a claims form, routes a ticket, or approves a vendor invoice has immediate, real-world consequences when it makes an error. The market for agents that actually take actions, reliably, is enormous. The market for agents that mostly take actions but sometimes hallucinate is much smaller, because most enterprises will not trust them at scale.
Scaled Cognition is positioning APT as the first model that clears the reliability bar needed for the second market.
The $600 Billion BPO Opportunity
Customer experience is Scaled Cognition’s entry point, but the stated target is the broader business process outsourcing market, which the company sizes at $600 billion.
That number is plausible. Global BPO revenues, covering customer service, IT support, HR operations, finance and accounting, and supply chain management, consistently rank in the $500-700 billion range across industry estimates. The thesis is straightforward: enterprises currently outsource these functions to third-party managed service providers partly because building internal operations at scale is expensive, and partly because the work involves enough variability that software alone has not been reliable enough to replace human judgment.
AI changes both constraints. As Roth frames it, companies are “beginning to insource what they once outsourced, replacing third-party managed services with AI workforces they own and control.” Scaled Cognition is building the infrastructure to make that insourcing trustworthy.
The near-term evidence supports the ambition. Companies using APT are on track to automate more than one billion customer service interactions over the next twelve months. That is not a projection about a hypothetical future product. It is a claim about deployments already running in production.
The unit economics work in the same direction. APT is described as smaller, faster, and less expensive than frontier models. At one billion customer service interactions per year, the gap between a reliable model and a frontier model with guardrails translates directly into margin. For enterprises in financial services or insurance where a single wrong interaction can trigger a compliance event, it also translates into risk reduction that is difficult to put a precise number on but easy to justify to a CFO.
Who Built This and Why It Matters
Dan Roth and Dan Klein previously co-founded and sold one of the first agentic AI companies to Microsoft. That history is relevant: they spent years inside enterprise AI deployment before starting Scaled Cognition. The reliability problem Roth describes is not academic. It is a problem they hit in production, repeatedly, before deciding the only real solution was a different architecture rather than better guardrails.
Dan Klein’s dual role as CTO and UC Berkeley professor of NLP brings the research depth the architecture bet requires. The claim that Scaled Cognition has done for conversational AI what verifiable RL did for coding is an extraordinary one, and it demands credibility beyond a funding announcement alone.
What This Means for Enterprise AI Teams
The reliability bottleneck is commercially validated. Khosla’s $100M, combined with Genesys deploying APT across 8,000+ enterprise customers before the round closed, confirms that frontier models are capable enough for most tasks but not yet reliable enough for high-stakes autonomous operations at scale. That bottleneck is real, it is attracting serious capital, and it will not resolve itself through prompt engineering.
VPC deployment changes the compliance calculus. Many enterprises blocked by data governance requirements rather than capability concerns will find APT’s self-hosted deployment option directly addresses the constraint. If the model performs reliably and data never leaves the enterprise perimeter, the compliance conversation in financial services and healthcare changes significantly.
The right evaluation framework is “reliability per dollar,” not just capability per dollar. Teams evaluating models for agentic workloads should add reliability benchmarks on their own internal data to the scorecard alongside inference cost and benchmark performance.
The Scaled Cognition funding fits a broader pattern in Enera’s coverage of the enterprise AI deployment wave in 2026: the frontier capability race is real, but the practical deployment bottleneck is not capability. It is the gap between what models can do in controlled conditions and what they do reliably across millions of production interactions.
The Broader Architecture Race
Scaled Cognition’s bet lands in a specific competitive context. OpenAI’s o-series reasoning models spend more inference compute checking answers before returning them: effective for verifiable tasks like coding, but expensive at scale. Anthropic’s extended thinking improves calibration but does not structurally prevent hallucination on in-domain factual tasks. Together AI’s open-source inference platform (which raised $800M on July 1) takes yet another angle: make inference cheap enough to run multiple verification passes.
Scaled Cognition’s claim is that reliability should be a training property rather than an inference-time patch. If that holds at scale, it represents a genuinely different point on the reliability-cost curve: not more expensive per interaction, but architecturally less likely to produce confident wrong answers.
With one billion customer service interactions committed through Genesys and Fortune 500 deployments already in production, Scaled Cognition will have real production data on that question within the next twelve months.
| Metric | Detail |
|---|---|
| Funding | $100M Series A |
| Lead Investor | Khosla Ventures |
| Strategic Investor | Genesys (also a customer) |
| Founders | Dan Roth (CEO), Dan Klein (CTO, UC Berkeley) |
| Prior Exit | Sold first agentic AI company to Microsoft |
| Flagship Model | APT (Agentic Pretrained Transformer) |
| Deployment | VPC and self-hosted available |
| Target Market | $600B BPO, starting with customer experience |
| Production Traction | 1B+ customer service interactions projected (12 months) |
| Key Partner | Genesys Cloud (8,000+ enterprise orgs, 100+ countries) |
| Announced | July 1, 2026 |