When OpenAI or Google needs to train their next-generation AI models to reason like a lawyer, think like a physician, or analyze data like a seasoned banker, they face a fundamental challenge. The models need to learn from humans who can consistently outperform them. Finding those experts, and structuring their knowledge in ways that actually improve model performance, is harder than it sounds.
This is exactly where Mercor has carved out its position in the AI supply chain. And it's why the company reached out to David Nagtzaam, founder of Decode, to contribute to two of their most ambitious internal initiatives: Project Argentum and the AI Productivity Index, known internally as APEX.
The Quiet Infrastructure Behind Frontier AI
Mercor operates in a space most people never think about. Before ChatGPT can answer a complex legal question or GPT-5 can draft investment analysis, someone has to teach these models what "good" looks like. Not through code, but through examples. Thousands of carefully constructed problems, responses, and evaluations that push models beyond their current capabilities.
The company has grown rapidly on this premise. Backed by Felicis Ventures, Benchmark, and General Catalyst, Mercor reached a $10 billion valuation in late 2025 after raising $350 million in Series C funding. Their platform now connects over 30,000 domain experts with the world's leading AI laboratories. The work is substantial: Mercor pays out more than $1.5 million daily to contractors providing the specialized human intelligence that machines cannot yet replicate.
But raw expertise isn't enough. The challenge lies in translating professional knowledge into structured formats that actually improve AI reasoning. This requires people who understand both the domains themselves and how AI systems learn.
Why Mercor Tapped Decode
David Nagtzaam's background made him an unusual fit for this kind of work. Most domain experts know their field deeply but struggle to translate that knowledge into AI training contexts. Most AI specialists understand the technical requirements but lack real-world domain experience across multiple industries.
Nagtzaam brings both. With over 16 years consulting across legal, financial, healthcare, and technology sectors, combined with hands-on AI implementation experience at organizations including Google, he sits at an intersection that Mercor found valuable for Project Argentum and APEX.
Project Argentum focuses on developing training rubrics that teach models how to reason through complex professional scenarios. The goal isn't simply getting models to produce correct answers. It's training them to demonstrate the kind of structured thinking that distinguishes junior analysts from senior practitioners. How does a skilled consultant break down an ambiguous business problem? What makes a physician's diagnostic reasoning reliable rather than lucky?
APEX takes this further, creating benchmark evaluations that measure how well frontier models perform economically meaningful tasks across investment banking, management consulting, law, and primary care medicine. Unlike academic benchmarks that test abstract capabilities, APEX evaluates whether models can actually do the work that professionals get paid for.
Nagtzaam contributed to both initiatives, developing evaluation frameworks and training scenarios drawn from his cross-industry experience. His work helped establish rubrics for domains where expert judgment involves weighing competing factors rather than finding single correct answers.
The Stakes for AI Development
Understanding why this work matters requires grasping a counterintuitive fact about AI improvement. To train GPT-5, OpenAI needs GPT-4 to fail. Model development depends on finding problems that current systems cannot solve but human experts can. These failure cases become the training data for the next generation.
As models improve, finding humans who can consistently stump them becomes harder. The pool of useful training data shrinks with each generation. This makes Mercor's expert network increasingly valuable, and makes the frameworks that structure expert knowledge increasingly critical.
The domains Nagtzaam worked on carry particular weight. Legal reasoning, medical diagnosis, financial analysis, and strategic consulting all involve judgment calls that resist simple automation. They require balancing competing values, reading between lines, and knowing when standard approaches don't apply. Teaching AI systems to navigate these subtleties determines whether future models become genuine professional tools or remain sophisticated pattern matchers.
What This Signals for AI Advisory
For Nagtzaam, the Mercor partnership represents a particular kind of validation. Consultants often face skepticism about whether their expertise translates beyond traditional advisory contexts. Working directly on the infrastructure that powers frontier AI development demonstrates something concrete about the applicability of cross-domain business knowledge to emerging technology challenges.
"The work we did on Argentum and APEX wasn't about teaching AI systems facts," Nagtzaam notes. "It was about encoding the judgment that comes from years of professional practice. How senior practitioners actually think through problems, not just the conclusions they reach."
This distinction matters as AI capabilities expand. Organizations looking to implement AI effectively need partners who understand both what these systems can do and the professional domains where they'll be applied. The Mercor collaboration positions Decode at that intersection for clients navigating their own AI adoption challenges.
Looking to transform your organization through AI and intelligent automation? Decode helps ambitious companies move from AI curiosity to measurable capability. Contact us to discuss how we can accelerate your AI strategy.