Ramp
Corporate cards and spend management with AI expense review

Mercor connects domain experts with paid AI training and evaluation work while helping enterprises build industry-specific benchmarks and source human data. This review covers its expert marketplace, APEX benchmark suite, and enterprise services.
Mercor is a platform that pairs domain experts with paid AI training and evaluation projects while giving enterprises a way to source human data and build benchmarks for their AI systems. It sits at the intersection of a talent marketplace and a research operation, and its public materials treat those two functions as parts of the same business.
On the talent side, Mercor recruits professionals across a wide range of fields and matches them to projects using AI that assesses their abilities. Listings on the site show the range of work available: a Sales and Marketing Expert role at $60–$70 per hour, a Chemicals Safety for Redteaming position at $65–$75, and a Retail Trader role at $400. Each listing notes how many people were recently hired, which gives applicants a rough sense of demand. The company reports an average contracted rate of $91 per hour, more than 697,000 roles created, and over $4 million in daily payouts. Those figures come from Mercor itself and describe the scale it claims, not an independent audit.
Mercor's research arm produces benchmarks in collaboration with established companies, and the approach is notably concrete. Rather than testing models on abstract puzzles, these evaluations use tasks drawn from actual professional work. The legal benchmark, Harvey LAB, was built with Harvey and contains 1,749 real legal tasks across 24 practice areas. Grading focuses on finished work products such as memos and redlines, which puts the emphasis on output quality rather than intermediate steps. The benchmark is open source, and its code is published on GitHub.
APEX-Accounting, developed with Ramp, measures whether AI agents can complete professional accounting work across accounting software, spreadsheets, and PDFs. The tasks are long-horizon and cross-application, meaning an agent has to carry context from one tool to another. APEX-SWE, built with Cognition, evaluates software engineering performance on tasks ranging from bug fixes to feature builds. Both benchmarks publish leaderboards, and supporting papers are available for readers who want to examine methodology.
Public leaderboards list model scores with margins of error, which is a useful detail for anyone comparing systems. On the Harvey LAB leaderboard, for instance, the top listed result is 27.4% with a margin of plus or minus 5.4%, and scores decline from there. The accounting leaderboard shows much higher percentages, with the leading entry at 71.5%. The gap between the two illustrates why benchmark choice matters: a model that looks strong on accounting tasks may struggle with legal work. Readers should treat any single score as evidence about a specific task set, not a general measure of capability.
For enterprises, Mercor offers services around human data, custom evaluations, and data monetization. The company works with clients to design benchmarks that reflect their own industries, products, and workflows, and it also helps businesses sell or license their data. Mercor's documentation argues that human input remains essential for pushing model capabilities beyond the current frontier, for building new capabilities, and for measuring system performance, while synthetic data is better suited to catching up with leading models. That position is consistent with the company's commercial interest in human labor, though the underlying argument about model self-training has support in published research.
The platform is most relevant to organizations that need expert-labeled data or realistic evaluations and to professionals with specialized knowledge who want project-based income. Its benchmarks have become a reference point for comparing AI agents on professional work, and the open-source releases make the methodology available for scrutiny. The main limitation is that results depend heavily on the specific tasks chosen, so benchmark scores should inform decisions rather than settle them.
Subscription pricing is not listed. Expert listings show hourly rates, such as $60–$70/hr for a Sales and Marketing Expert, $65–$75 for Chemicals Safety for Redteaming, and $400 for a Retail Trader role. The site also cites an average contracted rate of $91/hr. Prices are subject to change.
Features, pricing and availability may change. Always verify details on the official website. Maintained by steven · Last updated Oct 14, 2026 →
Corporate cards and spend management with AI expense review
Rippling is a workforce management platform that combines HR, IT, payroll, and finance tools on a shared data foundation, with AI features that draw on that combined data. It suits growing companies that want to replace scattered point solutions and spreadsheets with one system for managing people,
DataRobot is an enterprise platform for building, deploying, and governing AI agents, aimed at organizations that need agentic AI to run in production rather than in pilots.
AI assistant for elite law firms and in-house legal teams