Scaled Cognition
by Scaled Cognition
APT, a large action model trained to take actions instead of predicting text
Scaled Cognition is a model lab building APT, an Agentic Pretrained Transformer for customer-experience agents that must act on real stakes such as bank balances and insurance claims. It trains on synthetic agentic data with action-level objectives, claiming architecturally enforced policy compliance rather than prompt-based guardrails, and is small enough to self-host in a VPC.
Scaled Cognition is a Mountain View, California model lab building APT, an Agentic Pretrained Transformer — a large action model trained to predict actions rather than tokens, aimed squarely at customer-experience agents that must operate where the stakes are real: bank balances, medical records, flight changes and insurance claims. The company was founded by CEO Dan Roth, a former Microsoft corporate vice president of conversational AI, and CTO Dan Klein, a UC Berkeley AI professor and natural language processing researcher; the pair previously founded Semantic Machines, one of the first agentic AI companies, which Microsoft acquired in 2018, and the research team draws from CMU, Meta, AI2, MIT, Stanford, DeepMind, UC Berkeley and Amazon. APT-1, introduced in research preview in February 2025, was trained through a fully synthetic data pipeline combining simulation, reinforcement learning and agent-to-agent self-play, optimizing action-level rather than token-level objectives. Scaled Cognition claims this architecture makes policy compliance deterministic and structurally eliminates hallucination — permissions, constraints and business logic are enforced by the model rather than requested in a prompt — and says APT-1 outperforms existing foundation models on the tau-bench and ComplexFuncBench agentic benchmarks while remaining small enough to self-host inside an enterprise VPC or on premises. The platform surrounds the model with GenAPI simulation for pre-launch testing across thousands of scenarios, an Agent Evaluator that generates compliance reports, and production monitoring, all reachable over a RESTful API with no-code, low-code and SDK paths. In June 2026 the company raised a $100 million Series A led by Khosla Ventures with participation from Genesys, which had partnered with it in October 2025 and embeds APT-1 in Genesys Cloud for agentic virtual agent capabilities.
Heads of customer experience or contact center technology at banks, insurers, airlines and healthcare payers who need conversational agents to execute real transactions and cannot accept a nonzero hallucination rate on a policy-bound action.
Agents whose permissions and business rules are enforced by the model architecture rather than by prompt instructions, deployable inside the enterprise's own VPC or data center.
At a Glance
- Category
- AI Models & APIs
- Pricing
- Contact for pricing, Subscription
- Target Market
- CIOs, CTOs, Heads of Customer Experience, Contact Center Leaders, Enterprise Developers
- Deployment
- Hybrid, Self-hosted, API-based
- Headquarters
- Mountain View, United States
- Team Size
- 11-50
Key Features
- ✓Action-level pretraining
APT optimizes action prediction rather than next-token prediction, which is the basis of its determinism claim over general-purpose LLMs.
- ✓Architecturally enforced policy compliance
Permissions, constraints and business logic are enforced by the model rather than requested in a prompt, so they cannot be talked around.
- ✓Self-hostable footprint
Small enough to run inside an enterprise VPC or on premises, keeping regulated conversation data out of a third-party API.
- ✓GenAPI simulation testing
Exercises an agent across thousands of generated scenarios before launch and reports correctness scores and policy compliance gaps.
- ✓Agent Evaluator and monitoring
Produces compliance reports and watches production behaviour continuously, which is what an audited contact center actually needs.
- ✓Open integration surface
Standard RESTful API with no-code, low-code and SDK paths, and prompts and policies remain with the customer.
- ✓Low-latency serving
Reported sub-120ms time to first token and a 40% end-to-end latency reduction, which matters for real-time voice.
Capabilities
Use Cases
- •Banking account servicing
An agent checks balances and moves money while permission boundaries are enforced by the model rather than by prompt guardrails.
- •Insurance claims handling
Conversational intake and status updates on claims where every response must stay inside documented policy and coverage rules.
- •Airline itinerary changes
Agents rebook and modify flights, an action-heavy task where a hallucinated fare rule creates real financial liability.
- •Genesys Cloud virtual agents
Enterprises on Genesys Cloud use APT-1 to power agentic virtual agent capabilities through the existing platform partnership.
- •Pre-launch agent certification
Compliance teams run GenAPI simulations across thousands of scenarios to produce evidence before an agent touches customers.
Ideal For
Best For
- ✓Regulated customer service where an agent modifies account state — payments, claims, bookings — and a wrong action has legal consequences
- ✓Enterprises that cannot send customer conversations to a third-party API and need the model self-hosted in their VPC or on premises
- ✓Contact center teams replacing rigid IVR and decision-tree bots with conversational agents that still respect hard policy constraints
- ✓Organizations already on Genesys Cloud, where APT-1 is available through the existing platform partnership
- ✓Latency-sensitive voice and chat deployments, where the vendor reports sub-120ms time to first token in production
Not Ideal For
- ✗General-purpose assistant, coding, content or research workloads — APT is explicitly targeted at conversational CX, with other verticals only planned
- ✗Teams that need published, reproducible benchmark figures before buying, since the company states it tops tau-bench and ComplexFuncBench without releasing the numbers
- ✗Buyers wanting a broad reference list, as Genesys is effectively the only publicly named deployment and is also an investor
- ✗Self-serve or small-team adoption: there is no public pricing, no free tier and no open-weights release
- ✗Anyone needing frontier general reasoning, since the model is deliberately small enough to self-host and optimized for action prediction
Integrations
Deployment
Market Analysis
Pros
- ✓Unusually strong founding team — Dan Roth and Dan Klein previously built Semantic Machines, acquired by Microsoft in 2018
- ✓A genuinely different technical bet: action-level pretraining on synthetic agentic data rather than tool-use fine-tuning of a general LLM
- ✓Self-hostable in a customer VPC or on premises, which unlocks regulated buyers that cannot use a hosted frontier API
- ✓Validated by a large incumbent — Genesys both partnered and invested, and embeds APT-1 in Genesys Cloud
- ✓Production latency is real and measured: sub-120ms time to first token with a 40% end-to-end reduction on its serving stack
Cons
- ✗The company claims APT-1 tops tau-bench and ComplexFuncBench but publishes no numbers and names no competitor models, so the headline claim is unverifiable
- ✗'Eliminates hallucinations' is an absolute architectural claim with no independent third-party evaluation behind it
- ✗Scope is narrow by design — conversational CX only, with other verticals described as planned rather than shipped
- ✗Genesys is effectively the sole publicly named customer and is also an investor, so the reference list is thin and not arm's length
- ✗No public pricing, no trial, no open weights and an 11-50 person headcount, which is meaningful vendor concentration risk for a core CX system
- ✗No G2, Capterra or TrustRadius presence, so there is no independent user sentiment to read at all
Pricing
Enterprise
Contact for pricing
- ✓APT model access via RESTful API
- ✓VPC or on-premises deployment
- ✓GenAPI simulation
- ✓Agent Evaluator
- ✓Production monitoring
No pricing is published anywhere on the site and there is no free tier, trial or open-weights release; access is an enterprise sales conversation, and self-hosting in a customer VPC or data center implies a licence rather than per-token metering. Buyers already on Genesys Cloud may reach APT-1 through that platform's commercial terms instead of contracting directly.
Security & Compliance
Sources
This page was written from 4 sources, 2 on domains other than scaledcognition.com.
Stay Ahead of the Curve
Weekly enterprise AI insights for technology leaders. No spam, no vendor pitches—unsubscribe anytime.
SubscribeRelated Products
TwelveLabs
Video intelligence API that makes every hour of enterprise footage searchable, analyzable and agent-ready.
Engram
A learned memory layer that makes AI actually know your organization — at up to 100x fewer tokens.
Mistral OCR 4
Structure-aware document AI that returns bounding boxes, typed blocks, and per-word confidence scores.
Hugging Face
The AI community building the future.