Mercor’s expert-feedback engine scales to $500M run rate and $10B+ valuation

Mercor has turned specialist human feedback into a routinized supply chain for model refinement, and the numbers show it: the firm reports paying over $1,000,000 per day to contributors, moved from a $1M run rate to roughly $500M in under 24 months, and is estimated at more than $10 billion in value. That acceleration signals investors view expert scoring and task-specific evaluation as a structural layer for production AI, not a temporary cost center.

The company aggregates tens of thousands of domain specialists—spanning medicine, law, finance, engineering, comedy and more—to produce labelled judgments and rubrics used in reinforcement learning from human feedback (RLHF) pipelines. Large labs such as OpenAI, Google and Anthropic rely on similar human-in-the-loop stages; Scale AI and other service providers supply adjacent capacity, making a nascent market architecture with multiple suppliers and high strategic value.

Practically, experts are enlisted to grade model outputs against custom rubrics and to author corrective responses; firms use those labels to optimize reward models and fine-tune policy networks. Compensation ranges up to several hundred dollars an hour for niche work, creating a gig-style labor market that monetizes professional judgment but also raises questions about the long-term composition of skilled work.

In healthcare applications, physicians train models on layered scenarios that mix patient-facing prompts and clinician-level jargon, which materially reduces hallucination risk and improves triage precision when integrated into clinical decision support. Yet experts interviewed stress models remain assistive: they reduce administrative load and surface differential diagnoses, but do not replicate tacit clinical judgement obtained from in-person evaluation.

The business implications are twofold: first, investor bets on companies like Mercor imply recurring revenue potential from continuous model maintenance and localization tasks; second, workflows that require specialized rubrics—comedy, legal reasoning, and region-specific content moderation—create defensible demand for curated human feedback. That dynamic favors platforms that can both recruit deep specialists and translate qualitative judgements into quantitative reward signals.

Operational risks persist. Subjective tasks scale poorly because inter-rater variance undermines consistent reward shaping, and localization demands increase labeling complexity for humor or culturally specific content. Competitors such as Surge AI, Handshake and Micro1, plus large strategic players, create pricing pressure that could compress margins if commoditization occurs.