If AI is automating experts, why are AI companies buying more expert judgment?

AI is automating parts of legal review, software development and medical analysis. Yet the market for high-quality AI data is creating demand in the opposite direction.

Reuters reported on September 22 that Snorkel AI raised $350 million at a $3.5 billion valuation. Annualized revenue has risen above $350 million from about $20 million a year earlier.

Reuters said much of that growth came from a Data-as-a-Service business launched in September 2025, where Snorkel supplies finished datasets and reinforcement-learning environments rather than simply software.

The bottleneck is moving from labor volume to difficult judgment

Traditional AI-data work is often associated with repetitive labeling at scale. Frontier models change the problem because useful training and evaluation tasks have to become harder as the models improve.

Insurance underwriting, legal interpretation, complex coding failures and clinical reasoning require people who understand the domain well enough to design and validate difficult tasks.

Snorkel's Expert Community now spans 1,000+ domains across coding, STEM, health and medicine, finance and business, law, humanities and social sciences.

Hard problems require a system for validating expert judgment

Expert tasks rarely have one trivial ground truth. Experts can disagree, prompts can be ambiguous and a correct answer can still rely on flawed reasoning.

Snorkel says its quality process uses calibration, multi-stage and dual-expert review, consensus mechanisms, statistical sampling, provenance and audit trails.

The problem therefore becomes not just finding more people, but deciding which experts to use, which tasks matter and how conflicting judgment should be resolved.

Enterprise Environments turn company work into something agents can train on

Snorkel's Enterprise Environments simulate real company workflows rather than isolated question-answer tasks. Agents may need to read policies, inspect records, call internal tools, ask simulated users for missing information and update system state.

Its insurance work uses fictional companies, customer records, underwriting guidelines and imperfect information to recreate a multi-step underwriting workflow.

That moves the product away from simple labels toward a trainable representation of how expert work is actually performed.

The product unit shifts from human hours to packaged judgment

In a conventional professional service, revenue often scales with people and hours.

In the emerging data-product model, expert judgment can be transformed through task design, validation and quality control into reusable datasets, benchmarks and training environments.

The pattern is Expert Judgment → Packaged Intelligence.

Snorkel's current hiring reflects evaluation, quality and frontier benchmarks

Current roles include Director, Research - Evaluation & Training, Research Scientist - Frontier Benchmarks, Senior Product Manager - Data & Quality, Staff Product Manager - Data Products and Director, Forward Deployed Researcher - Data as a Service.

Its Data-as-a-Service organization also lists forward-deployed engineering, synthetic-data-generation and strategic-delivery roles.

The organizational signal is that competitive advantage is not only about producing more data, but about identifying the data and environments that expose model limitations and map to real customer work.

This does not mean every expert automatically becomes more valuable

Snorkel's Expert Community is largely project-based independent-contractor work. The economic value captured by the finished data product and the compensation paid to individual experts are not the same thing.

One company's growth also does not prove that wages or employment for every profession will rise.

The narrower claim is that a new market is buying expert judgment and converting it into AI training and evaluation assets.

BANSEOG VIEW | AI may reprice judgment more than manual effort

If early AI data markets rewarded the scale of repetitive human work, frontier AI can shift the bottleneck toward people who know what the model still gets wrong.

Professional expertise can create value outside the professional's primary employer by becoming training data, benchmarks, evaluations or simulated environments.

The useful talent-market question may become: is this expertise valuable only inside its current industry, or is a new market emerging that buys the judgment itself?

BANSEOG VIEW

Banseog View — Expert Judgment → Packaged Intelligence

Snorkel's growth is evidence of a market moving beyond commodity labeling toward packaged expert judgment.

As models improve, human data work can become more domain-intensive rather than simply disappear.

Expert judgment can migrate from billable hours into reusable datasets, benchmarks and enterprise environments.

SOURCES

Primary sources and references

Snorkel's growth does not establish that compensation or employment rises for all experts. Its Expert Community is project-based contractor work, and product value captured by the company can differ from individual expert compensation. ‘Expert Judgment → Packaged Intelligence’ is Banseog's analytical framing.