Why the next AI data race will be for human expertise and knowledge

In 2026, Thinking Machines Lab and Bridgewater Associates wanted to test whether artificial intelligence (AI) could help investors filter the large volume of news and research they review each day.

The resulting study tested six information-filtering tasks, including deciding whether an article was relevant. Bridgewater initially used examples labelled by non-experts to indicate what counted as relevant or useful, but many of those judgements proved wrong. Disputed examples were sent to Bridgewater investors for review and that feedback was used to fine-tune a specialized model.

The resulting model outperformed other models Bridgewater had tested, while also costing substantially less to run.

The experiment was narrow, but its broader lesson is simple: Financial information was readily available to feed into AI tools, but reliable judgement about which information mattered was harder to obtain.

Large language models (LLMs) are a type of foundation model, which are general-purpose AI systems that can be adapted to many different tasks. The most capable models available are often described as frontier models.

LLMs can process financial reports, legal documents and scientific papers at enormous scale. But access to that material does not reproduce the judgement of someone who has spent years working in the field.

Experts learn which details deserve attention, which mistakes matter and when the usual process is likely to fail. Some of these lessons can be documented, but others become apparent through the decisions people make. For example, an investment analysis can contain accurate information while focusing on the wrong signals. Producing more examples only goes so far without a dependable way to judge their quality.

Even then, expert judgement is not a single objective truth. Professionals can disagree because they have different objectives, risk tolerances or institutional priorities. A model trained on Bridgewater’s decisions learns what Bridgewater considers relevant rather than some universal form of investment judgement.

It’s this institution-specific judgement that competitors may struggle to reproduce when developing new AI tools such as foundation models, making this type of data crucial to an organization’s competitive advantage in the AI era.

Demand for specialized human input is already supporting a substantial commercial market. A fast-growing segment of companies now connects AI developers with doctors, lawyers, scientists and other specialists who can create and evaluate professional tasks.

One of these companies, Micro1, saw its gross annual run rate increase from $100 million to $500 million in eight months, according to a TechCrunch report. It also put similar firms Mercor at roughly $2 billion and Handshake at around $1 billion in gross annualized revenue. These figures indicate the scale of spending on specialist AI data and evaluation today.

The work of these companies goes beyond conventional data labelling. Mercor describes projects involving professional benchmarks, datasets and evaluation environments. It asks specialists to write realistic problems, review model outputs or spot mistakes that require domain knowledge.

This market could make generic professional expertise easier to acquire. But knowledge tied to one organization – information about its customers, previous decisions, internal processes and unusual edge cases – is harder to buy.

Much of the information behind good company decisions has never been published online. Organizations’ internal systems contain unusual customer cases, operational failures and decisions accumulated over years. Experienced staff know which exceptions deserve attention and when the normal process should not be followed.

That’s why law firm Kirkland & Ellis has committed $500 million to developing its own AI platform, with input from around 250 lawyers and more than 180 technology professionals, while continuing to use external AI products, according to the Financial Times.

As capable foundation models become available to more organizations, access may explain less of the difference between what companies achieve with AI. The difference could instead come from better retrieval, richer examples, stronger evaluations or better records of how previous decisions turned out. And so, a system that continually produces organization-specific feedback may be harder to copy.

When employees review AI outputs, they correct recommendations, reject plausible answers, escalate unusual cases and later observe whether decisions have been successful. Each intervention can reveal something about what good performance looks like.

Many organizations record the final decision, but the corrections that shaped it are harder to reuse. Recording those corrections and decisions could create useful material for future evaluations, retrieval systems or model improvements.

Synthetic data, which is artificially generated rather than collected directly from real-world observations, could reduce the need for some human-generated examples.

The World Economic Forum has already examined how synthetic data can ease constraints on AI development. And the 2026 International AI Safety Report, commissioned by the UK government, highlights mathematics, programming and formal reasoning as areas where synthetic approaches can work particularly well because answers can be checked against relatively clear standards.

Verification of synthetic data does not always have to come from a human either. Software tools, simulations, observed outcomes or other AI systems can provide feedback. Better automated evaluators could also reduce the need for expensive expert reviews.

The economically important question is how cheaply an AI-driven system can obtain trustworthy feedback about whether its output is good. Where verification can be reliably automated, cheaper synthetic data could become abundant. Where quality depends on scarce expertise or outcomes that take time to observe, good feedback may remain expensive.

Capturing judgement also creates governance questions. Employee corrections, customer interactions and professional decisions can contain confidential or personal information. Organizations need appropriate rights to reuse this data and must create controls around access.

As more organizations gain access to more capable foundation models, some of their competitive advantage may shift towards the feedback surrounding these models, including which outputs succeeded, which failed and which plausible answers experienced professionals rejected.

Some of that knowledge can be purchased or imitated, but a continuously updated record of how one organization makes difficult decisions and learns from the outcomes may be harder to reproduce. The next AI data race may therefore be less about collecting more information than obtaining reliable evidence about what good performance actually looks like.