Building Competitive Advantage Through Proprietary Workflow Data

Mercor CPO Osvald Nitski believes the current AI ROI problem stems from a lack of organizational patience and poor workflow integration rather than any failure of the technology itself. While the market focuses on model benchmarks, Nitski argues that real competitive advantage comes from long horizon tasks. These are complex, multi step processes that most enterprises have yet to automate. The result is that the primary bottleneck to AI success is shifting from model performance to high fidelity human data and specialized operational expertise. For leaders and founders, the path is clear: stop chasing generic benchmarks and start investing in proprietary, workflow specific data. Those who handle the operational friction of building these bespoke evaluation systems now will create a durable advantage that competitors waiting for plug and play solutions will struggle to overcome.

The Myth of the 90-10 Split

Conventional wisdom suggests that 90% of enterprise workflows will soon be handled by commoditized open source models, leaving only a 10% niche for frontier providers. Nitski dismisses this as a calculation based on limited, existing demand. He argues that most companies are not even attempting the high value, long horizon tasks like fully automated procurement or complex legal reasoning that define the true frontier.

"There is a whole category of latent demand that people are not even, these are things that people are not even trying to do with models yet."

-- Osvald Nitski

When you look ahead, that 10% is not a shrinking slice; it is an expanding frontier. Organizations that treat AI as a binary tool will be left behind by those who treat it as a continuous, uncapped process of improvement.

Why Flexible Tools Create Operational Chaos

In the rush to capture AI demand, product teams often build for maximum flexibility, trying to support every customer request. Nitski reflects on Mercor’s own experience, noting that building a tool to support hundreds of different annotation workflows created massive internal friction.

The downstream effect was a loss of focus. By trying to be everything to everyone, the product team increased their own operational overhead. The lesson is counter intuitive: immediate responsiveness to customer requests can create chaos that prevents long term scaling. Nitski’s team had to pivot to imposing guardrails, prioritizing only those workflows with enduring demand. The advantage here is not in the speed of feature delivery, but in the discipline to say no to non scalable requests.

The Hidden Cost of AI-First Talent

Nitski points to a systemic imbalance: the concentration of AI deployment expertise in a small group of San Francisco based engineers. This creates a temporary services boom where companies like Palantir and Microsoft are selling the knowledge of how to deploy, not just the software.

"I think it is the future for the short term as the knowledge of how to use AI gets disseminated throughout industry... eventually it will be a job function similar to software engineering."

-- Osvald Nitski

The consequence of this knowledge gap is that enterprises are paying a premium for forward deployed engineers. Over the next 12 to 18 months, as this expertise spreads, the competitive advantage will shift from those who have the engineers to those who have successfully integrated AI into their internal culture. If you are hiring today, Nitski suggests biasing toward senior talent who prioritize business impact over tool fluency, as the skill of using AI tools is rapidly commoditizing.

The Robotics Waymo Moment

While the market looks for a ChatGPT moment in robotics, Nitski suggests a different trajectory. Drawing a parallel to the evolution of autonomous vehicles, he argues that robotics will likely follow a path of slow, localized deployment like Waymo in San Francisco rather than a sudden, global inflection. The implication for investors and builders is that the physical data market will be the next major frontier. Because physical environments are harder to simulate and scale than software, the first companies to master high fidelity, environment based training data will hold a significant, durable advantage.

Key Action Items

• Audit your Long-Horizon Tasks: Identify workflows that currently require weeks of human oversight. Over the next quarter, stop asking if a model can do the task and start mapping the data required to improve its performance continuously.
• Implement Enduring Demand Guardrails: Review your product roadmap. If you are supporting workflows that do not scale, prune them now. The discomfort of telling customers no today is a prerequisite for the efficiency you will need in 12 to 18 months.
• Shift Hiring Focus to Judgment: Stop testing for AI tool fluency in interviews. Use whiteboarding sessions to test for systems design and experimental rigor. The ability to judge what to automate is becoming more valuable than the ability to write a prompt.
• Invest in Proprietary Data Moats: If you are a founder, stop relying on public benchmarks. Build internal eval sets that reflect your specific business outcomes. This pays off in 12 to 18 months as your model becomes uniquely tuned to your operational goals.
• Prepare for the Robotics Shift: If you are in a capital intensive industry, begin exploring how world state data, or simulated environments, might impact your operations. This is a three year investment horizon, but it is where the next major data demand is forming.