Harvey's $15.5B Vertical Neo-Lab: Beyond Legal ChatGPT to Closed-Loop AI Training

This article is part of a series examining enterprise AI deployment benchmarks, starting from Sequoia Capital's "sovereign AI" concept to dissecting application-focused Neo-Labs.

Legal AI company Harvey is often categorized as a vertical SaaS for legal Q&A and contract review, but that only covers its user-facing product layer. As of September 2026, Harvey's disclosed valuation is $15.5 billion , with over 2,400 customers across 70+ countries and a team exceeding 1,200 people . More importantly, the company connects legal workflows, professional personnel, evaluation systems, agent runtime frameworks, and model training into a single system.

The term Vertical Neo-Lab (New Lab) is an analytical framework, not Harvey's official classification. It refers to AI companies that deeply penetrate a specific industry and continuously link task definition, delivery, evaluation, and training.

I. Evolution Sequence: Training the Model Comes Last

A common vertical AI development path is:

Find open-source model → Fine-tune an "industry LLM" → Package as enterprise brain demo → Hunt for paying customers

Harvey's publicly visible sequence starts from real legal tasks:

Enter real legal workflows
↓
Legal Engineer goes on-site (make implicit business explicit)
↓
Define Eval & Rubric (clarify what counts as qualified)
↓
Build Harness & Cloud Runtime (handle scheduling, security, execution)
↓
Construct synthetic training environment driven by scoring criteria
↓
Launch post-training, train Tenet

This order solves how tasks are executed and accepted before deciding what the model needs to learn. Model weights are one link; task definitions, evaluation standards, and runtime environments equally participate in final delivery.

II. Two Roles That Turn Legal Experience into Engineering Assets

Two roles handle the translation between professional experience and product engineering.

1. Legal Engineer: On-Site Legal Engineers

Who they are: Hybrid personnel with both legal business and technical implementation skills.

What they do: Refine scenarios around clients' real documents and tasks, configure prompts, multi-step Agent workflows, and participate in delivery training.

Value: They bring back clients' actual tasks, constraints, and failure modes into the product and engineering system.

2. ALR (Applied Legal Research): Turning Professional Standards into Evaluations

Senior legal practitioners and AI researchers convert professional judgment into Benchmarks, Datasets, Rubrics, and feedback signals for training. This lets engineering answer two concrete questions: what makes a legal deliverable qualified, and where does the model err across jurisdictions, clauses, and factual relationships.

III. Training Domain Models Without Customer Data

Harvey's Tenet research notes explicitly state that no customer data was used in this post-training. Training material consists of synthetic data, public legal data, and human expert data. Crucially, scoring criteria are established before training tasks, forming a synthetic environment production line:

Define tasks and scoring criteria: Domain experts write complex task scenarios, acceptance standards, and easily missed risk points.

Generate virtual legal materials: Construct contracts, emails, memos, and other synthetic materials around the task.

Run long-horizon tasks: Let the model read, retrieve, call tools, and generate deliverables within the environment.

Automated and expert scoring: Use verifiers and expert rubrics to feed results back for post-training.

The Harvey Tenet model, revealed in August 2026, uses Kimi K3 as its base and was trained in collaboration with Fireworks Research. Official disclosures show approximately 1,750 task environments were constructed, over 10,000 training trajectories run, using about 150 NVIDIA B300 GPUs over roughly two months .

IV. Decomposing Public Capabilities into Six Layers

From Harvey's product and Tenet research materials, its public capabilities can be organized into six layers:

Application Interaction Layer: Spaces, Word/Outlook plugins, and Matter collaboration entry points.

Domain Knowledge & Memory Layer: Vault, legal reference materials, and the firm's own work standards.

Scheduling & Runtime Layer: Agent framework, tool calling, sandbox environment, and model routing.

Model Layer: General closed-source models, open-weight models, and Tenet.

Evaluation & Learning Layer: LAB legal agent benchmark, expert rubrics, and synthetic task environments.

Enterprise Governance Layer: Data retention, data residency, and access isolation controls.

A legal task enters via the application layer, calls institutional knowledge and tools, is orchestrated by the runtime framework for model execution, and results are checked against evaluation standards. Synthetic task environments and expert feedback turn failure cases into subsequent training material. These six layers are not an official fixed architecture diagram from Harvey, but a framework for understanding how its product, delivery, and training connect.

V. Five Takeaways for Enterprise AI Founders

Harvey's approach translates into five verifiable startup choices:

Prioritize high-value vertical workflows: Scenarios need sufficiently high delivery value, obtainable digital materials, and human experts who can review results — e.g., supply chain review, quote verification, technical compliance, or manufacturing root-cause analysis.

Build evaluations before discussing fine-tuning: Without a repeatable domain benchmark, it's hard to judge whether swapping prompts or models actually improves accuracy, completeness, and risk.

Involve industry experts in task engineering: People who understand industry rules and can configure Agent workflows can translate client language into tasks, tools, and acceptance criteria.

Distinguish context updates from model training: Context injection and Memory solve single-session or cross-session information use; weight training changes the model itself. Their data permissions, validation methods, and risks differ.

Configure early team around the closed loop: A viable small-team setup is:

1 Field Deployment Engineer (FDE) blending domain knowledge and on-site delivery + 2 Product/Agent Engineers + 1 Eval/Data Engineer + 1 Integration Engineer

The Harvey case illustrates that vertical AI competitiveness can reside in a continuously operating organizational closed loop — entering client tasks, defining acceptance standards, stabilizing execution, and converting usable feedback into the next round of improvement.

Sources: Harvey official company page, August 2026 Tenet research preview, and September 9, 2026 financing announcement.