The Evaluation Bottleneck: What Harvey's Model Strategy Reveals About Enterprise AI

1. The Signal Event: Harvey Moves In-House

In June 2026, Harvey—the dominant legal AI platform serving over 500 customers, including approximately 42% of the Am Law 100—made a profound strategic pivot. Valued at $11 billion following its March funding, the company announced it was building custom legal-specific foundation models.

This wasn't merely a fine-tuning exercise; Harvey detailed an agentic system orchestrating legal tools alongside frontier models, and signaled plans to open-source models, training data, and benchmarks through its newly formed Harvey Labs division.

The explicit ambition, as articulated by co-founder Gabe Pereyra, is to "create the foundations for law firms to build their own specialized models and own their own intelligence." This statement is pivotal. It marks the arrival of the in-house model shift in what is arguably the most conservative industry in the world. While Harvey had previously collaborated with OpenAI on custom case law models in 2024, this move represents a definitive step toward platform independence and vertical integration.

2. The Expert Data Economy

The moment vertical AI leaders move in-house on models, they encounter the data exhaustion problem. The open web has largely been scraped clean; what remains lacks the density and professional rigor required for specialized applications.

Consequently, frontier labs and vertical AI platforms are spending extraordinary sums to procure domain data directly from practitioners. The rise of the expert data economy illustrates the scale of this demand. Mercor, an expert-data marketplace matching AI labs with lawyers, doctors, and scientists, reached a ~$10B valuation in late 2025; Mercor counts major AI labs including OpenAI, Anthropic, Meta, and Google among its reported clients.

Similarly, data development platforms like Snorkel AI have become critical infrastructure. Snorkel served as Harvey's data partner for BigLaw Bench: Research, released in March 2026 — a benchmark designed to expose the hard case-law research problems where frontier models still fail.

The valuations of human-in-the-loop data vendors—Scale AI, reportedly valued around $29 billion following Meta's 2025 investment, and Surge AI, reported to have reached roughly $1.4 billion in bootstrapped annual revenue—underscore the fundamental truth: in the current paradigm, domain expertise must be purchased as structured data.

3. From Benchmarks to Rubrics

Procuring expert data, however, only solves the training input. The subsequent realization for firms building proprietary models is that evaluation is the actual binding constraint. Generic academic benchmarks (MMLU, GPQA) are effectively saturated; they no longer predict professional utility or agentic reliability.

This saturation is evident across domains. In August 2025, OpenEvidence scored a perfect 100% on the United States Medical Licensing Examination (USMLE). Yet, acing an exam benchmark provides little assurance regarding a model's ability to safely automate real clinical documentation or navigate complex differential diagnoses.

The industry is shifting toward expert-built, domain-specific evaluation. Harvey's release of Legal Agent Bench in May 2026—backed by Nvidia, OpenAI, Anthropic, Mistral, and DeepMind—is the archetype. It is an open-source evaluation suite comprising over 1,200 agent tasks across 24 practice areas, evaluated against more than 75,000 expert-written rubric criteria.

A critical finding from Harvey's research indicates that model answers become operationally unhelpful when they satisfy less than ~60% of a task's rubric criteria. This suggests that fine-grained, expert-defined evaluation—rather than holistic vibes-based grading—is what separates usable systems from liabilities. As AI development moves toward reinforcement learning environments and "agent gyms," training and evaluation are converging on the same scarce resource: realistic, domain-faithful task rubrics.

4. Anatomy of a Modern Benchmark

Examining the construction of serious, modern benchmarks reveals why they are so difficult to produce. OpenAI's HealthBench, released in May 2025, was built using 262 physicians from 60 countries to draft 5,000 realistic multi-turn health conversations. Each conversation was graded against custom rubrics authored by those same physicians.

In the financial sector, FinanceBench—developed by Patronus AI, Contextual AI, and Stanford—provides 10,231 questions on publicly traded companies, complete with answers and precise evidence strings designed for open-book financial QA. And Harvey's Legal Agent Bench, as noted, spans over 1,200 tasks with more than 75,000 individual rubric criteria.

Across these efforts, a common anatomy emerges:

1. Recruit credentialed domain experts (doctors, lawyers, financial analysts).
2. Author realistic, often synthetic, scenarios or tasks.
3. Write fine-grained grading rubrics for every single task.
4. Run models against these tasks and score their outputs according to the rubrics.
5. Refresh this dataset continuously as models evolve, data leaks, and regulations shift.

Crucially, the rubric—not just the task itself—is where the vast majority of expert labor and financial cost resides.

5. Finance and the "Effective Challenge"

While law and medicine are currently grappling with these evaluation challenges, the finance and insurance sectors are on an identical, arguably more formalized, trajectory.

In banking, the Federal Reserve's SR 11-7 guidance on model risk management has required independent validation of models since 2011. Regulators have actively begun applying its core principles to AI, ML, and generative AI. The guidance is currently being superseded by SR 26-02 to better reflect the AI-driven risk landscape. The foundational principle of this regulatory framework is that model validation must be independent of model development—a concept termed "effective challenge." As compliance professionals note, the team validating the model cannot report to the team that built it.

In the insurance sector, the NAIC's Model Bulletin on the Use of AI Systems by Insurers, issued in December 2023, has now been adopted by roughly half of U.S. states. The NAIC has also developed an AI Systems Evaluation Tool, reportedly being piloted across multiple states in 2026. Regulators are no longer just issuing guidelines; they are building their own evaluation instruments to audit enterprise systems.

The analytical takeaway is clear: banking is the one regulated industry possessing a mature, decades-old evaluation discipline. Its core tenet—independent, continuous validation by parties entirely separate from the development team—is exactly what AI evaluation in healthcare and legal currently lacks. Finance demonstrates exactly where the rest of the regulated economy is headed.

6. The Synthetic Substrate

If domain-faithful data is the prerequisite for rigorous evaluation, how do regulated industries acquire it when real matters are strictly confidential? The answer is synthetic generation.

Gartner has projected that synthetic data will come to overshadow real data in AI models. In regulated environments, it is often the only viable route. The HIPAA Privacy Rule, attorney-client privilege, and standard confidentiality agreements present an absolute bar to using real patient or client data for model training or evaluation.

The strongest evidence for this necessity lies in Harvey's own methodology. Despite having over 500 law firm customers and access to immense volumes of interaction data, Harvey builds its Legal Agent Bench tasks on synthetic matters—for example, M&A deals constructed entirely from synthesized, non-privileged documentation. If a $11 billion company with unparalleled market access must rely on synthetic data for evaluation, the synthetic substrate is undeniably the compliance-safe baseline for the industry.

7. The True Cost of Evaluation

The typical current practice for evaluating AI systems inside most enterprises consists of ad-hoc "vibe checks" by employees, small hand-built golden datasets that quickly grow stale, or relying on LLM-as-a-judge (models grading other models). The latter approach is notoriously unreliable without highly detailed, expert-written rubrics and is known to exhibit significant self-preference biases.

Moving beyond vibe checks requires interacting with a bifurcated expert labor market. While crowdwork data labeling can be sourced for a few dollars an hour, credentialed professionals—physicians, attorneys, and CPAs—authoring rubrics and grading outputs command far more. The contract domain specialist rate generally sits between $50–$100+ per hour, while highly specialized practitioners on platforms like Mercor can demand $85–$200+ per hour for complex professional tasks.

To back-of-the-envelope the cost honestly—and this is an illustrative estimate, not a sourced figure: if an effort on the scale of HealthBench (5,000 scenarios) required even a few physician-hours per scenario for drafting and rubric authoring, the total would run to thousands or tens of thousands of expert hours. At the professional rates cited above, that lands in the millions of dollars—before a single model is even graded. This excludes the operational overhead of recruiting, quality assurance, adjudicating disagreements among graders, and building the software infrastructure to manage it all.

More severely, this is an order-of-magnitude estimate for a single snapshot in time. Frontier models are updated on a cadence of weeks to months, regulations shift, and evaluation datasets inevitably leak into training corpora, rendering them stale. The evaluation spend is a recurring operational expense, not a one-time capital investment.

This leads directly to the first-party problem: building in-house evaluations means the same teams building the AI system are grading it. This creates the exact conflict of interest that the Federal Reserve's SR 11-7 was designed to prevent. Independent evaluation is fundamentally a governance requirement, not merely an engineering nicety.

For a hospital system, mid-size law firm, insurer, or biotech company, replicating even a fraction of this evaluation apparatus in-house is economically irrational. This realization is what makes external, continuous evaluation infrastructure the predictable next layer of the enterprise AI market.

8. The Infrastructure Gap

The progression proves the argument end-to-end: vertical leaders move in-house; they hit the expert data bottleneck; they recognize evaluation as the true binding constraint; they are forced to run that evaluation on synthetic data to maintain privilege; and they discover that doing it themselves is prohibitively expensive and fundamentally conflicted.

Harvey can afford to build this immense apparatus internally, buoyed by reported venture funding exceeding $1 billion. But what of the thousands of hospitals, biopharma companies, insurers, and law firms that intend to take Gabe Pereyra up on his offer to "build their own specialized models and own their own intelligence"?

These organizations inherit the data-evaluation-training triad without Harvey's resources. Few can justify the multi-million-dollar, recurring spend on expert networks that the arithmetic above implies, and fewer still have the internal engineering capacity to construct continuous, synthetic evaluation pipelines from scratch. This infrastructure gap—expert-grade, synthetic, independent, and continuous evaluation delivered as a service—is the structural opening in the enterprise AI market.