Scaled Cognition

What it is

Scaled Cognition is a model lab, not another agent vendor. It trains its own model for customer-service automation — APT, the Agentic Pretrained Transformer — and sells it with a platform for building, simulating, and monitoring agents on top. The argument is architectural: a general-purpose LLM is trained to predict tokens and then has tool use, guardrails, and policy checks bolted on, while APT is trained against action-level objectives — which API to call, which business rule applies, in what order — so policy adherence is a property of the model rather than of the prompt wrapped around it. The company markets APT as a model that “never guesses”; read that as the design target, not a guarantee you can skip testing for.

The founders have done this before. CEO Dan Roth and CTO Dan Klein, a UC Berkeley AI professor, previously co-founded Semantic Machines, which Microsoft acquired in 2018. Khosla Ventures led a $21M seed in 2023, and on 25 June 2026 led a $100M Series A with Genesys participating, at a reported valuation of about $750M (SiliconANGLE, citing the Wall Street Journal). APT-1 launched in research preview in February 2025, built by a team of around 20; the company says it outperformed frontier models on the Tau-Bench and ComplexFuncBench agentic benchmarks, but published no per-task scores in the launch post. The Series A release claims production deployments at Fortune 500 firms in financial services, healthcare, telecom, and insurance — without naming them.

Why it shows up in CX and support stacks

• It is the model underneath, not the agent on top. Most buyers meet APT inside Genesys Cloud. Genesys announced the partnership on 16 October 2025, invested, and on 10 February 2026 unveiled the Genesys Cloud Agentic Virtual Agent, powered by APT-1 as a “large action model,” with general availability set for Genesys’s fiscal Q1 2027 (February–April 2026). M&T Bank and Banco Pichincha were named as early evaluators. Genesys puts the reachable base at 8,000+ organizations.
• The simulation tooling is the practical differentiator. AgentTwin generates agents and test conversations from your historical customer interactions, GenAPI runs agents against thousands of simulated scenarios with mocked backend APIs, and an Agent Evaluator scores correctness and policy compliance across repeated runs before launch. Production monitoring then watches for drift. Teams that have been burned by an agent that passed a 50-conversation pilot and failed at 50,000 are buying this part as much as the model.
• It deploys where regulated buyers need it. APT runs on-premises, in your VPC, or hosted; access is a stateless REST API with no-code, low-code, and SDK options. Its inference partner Baseten reports time to first token under 120 ms — the number that decides whether a voice agent feels like a conversation or a hold queue.

Pricing reality

There is no public price list for direct deals; it is a sales-led enterprise contract. The one published meter is the Genesys route: Genesys documentation prices an Agentic Virtual Agent interaction at 1.2 AI Experience tokens, with tokens listed at USD 1.00 each — roughly $1.20 per interaction at list, before any enterprise discount, and on top of Genesys Cloud CX seats ($75–$240 per user per month by tier). On paper that undercuts Zendesk at $1.50 per automated resolution, but it is per interaction, not per outcome, so unresolved conversations still bill: at 70% containment the effective cost per resolved interaction is about $1.71, above Zendesk and well above Intercom Fin at $0.99 per outcome. For a direct deployment, expect a six-figure platform commitment comparable to Sierra or Parloa — an inference from the buyer profile, not a quoted number.

Best for

Contact-center and CX platform leaders at regulated enterprises — banks, insurers, telcos, healthcare payers, and the BPOs serving them — whose AI agent pilot failed on policy adherence rather than on conversation quality, and who either already run Genesys Cloud or need the model hosted inside their own VPC. The scoped use case: transactional flows where a wrong action is a compliance event — changing a payment plan, verifying identity before an account change, processing a claim step.

Skip it if your problem is FAQ deflection (any help-desk-native AI handles that), if you want a finished agent product your support ops team configures without engineering, or if you are not on Genesys and cannot staff a model-platform deployment.

Versus the alternatives

The comparison is not like-for-like: Scaled Cognition sells the layer that the agent vendors buy from OpenAI, Anthropic, and Google. Against the agent-layer leaders, pick Sierra when you want a vendor to own the outcome end to end and will accept outcome-based billing, and Decagon — the fastest-growing entrant in the segment — when your volume is chat-led and your ops team wants to write procedures in plain language. Pick Parloa for voice-first contact centers billing per minute. PolyAI is the closest structural peer, running its own Raven dialog model; choose it when voice is the whole problem and you want a managed service. Among the installed base, Salesforce Agentforce and Zendesk win when consolidation on the system you already pay for beats a better model. Choose Scaled Cognition when the failure you are fixing is the model going off-policy, and you would rather change the model than add another guardrail.

Watch-outs

• “No hallucinations” is a marketing claim, not a measured rate. The launch post asserts benchmark leadership without per-task scores, and no independent evaluation on customer-service workloads has been published. Guard: run your own 500-scenario policy-adherence test through AgentTwin and GenAPI against the frontier model your current vendor uses, and write the measured error rate into the contract as an acceptance criterion.
• Through Genesys, you are buying Genesys’s product, not Scaled Cognition’s. Roadmap, pricing, and support sit with Genesys, and the documentation we checked still listed agentic virtual agents as coming soon in some sections. Guard: confirm general availability in your region and tenant in writing, and ask which APT version your org runs and how model upgrades are rolled out and regression-tested.
• Per-interaction billing charges for failures too. At 1.2 tokens per interaction, a conversation that escalates to a human still consumes tokens. Guard: model cost per resolution as $1.20 divided by your containment rate, and set a containment floor below which the deployment is reviewed.
• Customer evidence is thin and unnamed. Beyond Genesys and its early evaluators, no production customers are public. Guard: ask for two reference calls in your industry before any multi-year commitment, and cap the first term at 12 months.