SC
Executive Brief — Scaled Cognition & the LAM Thesis
Competitive Intelligence · Strategic Assessment
Prepared for Morsy Cheikhrouhou
Date 2026-10-01 · Sources verified 2026-10-01

Large Action Models: Real Technology, Real Marketing, One Real Customer

Genesys just named Scaled Cognition's APT-1 the foundation of its next contact-center product. The startup raised a $100M Series A from Khosla Ventures in June 2026, has a serious research team led by a former Microsoft AI VP and a UC Berkeley professor, and claims its model "eliminates hallucinations" by training on actions rather than text. The technology is credible. The proof points are not yet independently verified. Watch for FY27 Q1 GA (calendar Feb–Apr 2027) — that's when we will see whether "LAM" is a category or a coat of paint.
$100M
Series A (Jun 2026)
39
Employees (+77% YoY)
3
Founders, repeat team
1
Named customer (Genesys)

What a LAM Actually Is

Genesys's Large Action Model is an AI trained to execute a sequence of approved steps inside a customer's systems, not to write a thoughtful paragraph about them. The model operates only on declared actions (authenticate, check availability, apply credit, close ticket) — it cannot invent a button. A separate "Process Reward Model" grades each action against the goal, and that score becomes the training signal. Compared with a typical chatbot, the difference is not intelligence — it is structure: the model has a rulebook and a checklist, and is graded on whether the work got done.

Under the hood, APT-1 is a transformer fine-tuned on enterprise action traces (API calls, ticket closures, refund flows) rather than web text. Genesys wraps it in its contact-center product and sells it under the name Cloud Agentic Virtual Agent. Scaled Cognition is paid per enterprise contract; the model itself is hosted-only — no public API pricing, no open weights, no Hugging Face presence.

What is new vs. what already existed. Tool-using AI has shipped since OpenAI's function-calling API in 2023, and LangChain / AutoGen / Salesforce xLAM have offered similar patterns for years. The genuinely new piece is the Process Reward Model and the "BRIDGE" training loop — grading the model on action-completion, not language plausibility. The rest is positioning.

The Company Behind It

Basics

  • Founded: 2023
  • HQ: Mountain View, CA (+ Amesbury, MA office)
  • Stage: Series A
  • Total disclosed funding: $100M+ (1 prior round undisclosed)
  • Valuation: Not disclosed
  • Headcount: ~39 (Sep 2026), +77% YoY
  • Public customers: 1 named (Genesys)

Founders

  • Dan Roth — Co-Founder & CEO. Ex-Microsoft CVP of Conversational AI. 4-time founder.
  • Dan Klein — Co-Founder & CTO. UC Berkeley professor of AI/NLP, former Microsoft Technical Fellow.
  • Damon Pender — Co-Founder & CFO.

All three previously co-founded Semantic Machines, a conversational-AI startup acquired by Microsoft in 2018. This is their second swing at the same problem, with deeper research and a more specialised focus.

Investors

  • Lead: Vinod Khosla / Khosla Ventures (announced 2026-06-25)
  • Strategic: Genesys (also a customer and product partner)
  • No other named investors in the Series A — depth of the round is Khosla + Genesys

Research Pedigree

  • Klein is one of the most cited NLP researchers alive
  • Roth led Microsoft's Conversational AI division at scale
  • Team: ex-Microsoft, Meta, DeepMind, Google, Allen Institute (AI2)
  • Academic roots: UC Berkeley, Stanford, MIT, CMU
  • Published at NeurIPS 2025 (booth T21)

What They Claim vs. What Is Verified

ClaimStatusSource
"Pass^100 reliability" — same task succeeds 100 of 100 runs Vendor-only scaledcognition.com product page
"Eliminates hallucinations" Half-true Independent researcher (Jun 2026): "no audited disclosure of architecture, parameters, or training data"
"Frontier CX model" Marketing No parameter count, training data size, or compute disclosed
"Industry's first autonomous agentic virtual agents" Disputed Salesforce Agentforce, ServiceNow, Microsoft Copilot Studio predate the claim
$100M Series A, Khosla-led, Genesys invested Verified Globe Newswire, 2026-06-25; confirmed by Genesys press release
39 employees, Mountain View + Amesbury Verified LinkedIn company page, Sep 2026
Named customers beyond Genesys None disclosed No public logo wall, no case studies, no conference references

What This Means for RingCentral

Strategic read

  • RingCentral's posture is "expose our data to other LLMs" (ChatGPT plugin, Claude MCP connectors, AIR Pro for healthcare). Ecosystem play, not owned-stack.
  • Genesys's posture is "own the action layer" via Scaled Cognition. Scaled Cognition becomes a strategic dependency for Genesys the same way OpenAI is for Microsoft.
  • Risk for RingCentral: if LAM-style action orchestration becomes the buyer-must-have, RingCentral can be relegated to the data layer. That's defensible short-term, vulnerable long-term.
  • Opportunity: Scaled Cognition is a thin layer today. RingCentral could license APT-1 (or a competitor's LAM), re-skin it, and ship the same wedge in 6–9 months.

Watch list — 2027

  • Feb–Apr 2027: Genesys Cloud Agentic Virtual Agent GA. First named customers, first measured outcomes (AHT, CSAT, resolution rate).
  • Independent benchmarks: a peer-reviewed or third-party eval of APT-1 vs. GPT-class models on multi-step enterprise tasks.
  • Competitive moves: Five9 / NICE / Webex / Zoom either license a LAM, build one, or partner with a different lab (Cohere, AI21, or another stealth startup).
  • RingCentral response: any acquisition, in-house build, or partnership announcement in the action-AI space.

Bottom line

Real technology, real team, one real customer, unverified claims. Treat "LAM" as a positioning word for action-trained AI that already exists in less branded forms. Do not treat APT-1 as a market force until either (a) Genesys publishes a paying-customer case study with measured metrics, or (b) an independent benchmark confirms the "Pass^100" claim. Earliest credible read: FY27 Q1 (Feb–Apr 2027).

Sources

Executive Brief · Bobby agent · 2026-10-01
Light theme · responsive · self-contained