Appen

Appen is an AI data services provider founded in Sydney in 1996 and listed on the Australian Securities Exchange; through a global crowd network and an annotation platform it delivers the human data that model builders need for training, fine-tuning and evaluation. It neither builds nor sells foundation models, and its stance is 'we do not compete with our customers, we only make them stronger' — only a neutral data vendor gets handed the full brief.

Core Capabilities

• ADAP annotation platform: unified annotation for text, image, audio, video, 3D point clouds and 4D, with multi-level review, gold test, AI pre-annotation and a REST API.
• CrowdGen crowd network: 1 million+ contributors, 200+ countries, 500+ languages and 100+ domains.
• Alignment and reasoning data: RLHF, SFT, red-team prompt sets and CoT reasoning traces; the company says it has delivered 250,000+ lines across 70+ dialects.
• Agent and physical AI: tool-call logs, golden trajectories and RL environment design, plus LiDAR and robot trajectories.
• Model evaluation and audit: third-party hallucination benchmarks, A/B preference tests and bias detection, alignable with the EU AI Act and the NIST AI RMF.
• Ready-made datasets: 597 OTS datasets with perpetual, non-exclusive commercial training rights, and enterprise data can be licensed after de-identification with ownership retained.

Best For Post-training and multilingual corpora, autonomous driving and embodied AI data, model audits in regulated industries, and frontier labs that require data neutrality. Small teams that only need a few hundred labelled items and want self-serve onboarding should not consider it.

Unique Advantages After Meta took a stake in Scale AI, customers began to fear that their data would leak, and Appen's no-models stance turned directly into orders; China revenue reached 102.9 million US dollars in FY25, up 75% year on year, and it is one of the few listed data vendors serving the US and Chinese sides at the same time.

Editorial Verdict Everest Group named it a Leader in the 2024 PEAK Matrix, and its compliance credentials and 30 years of craft are top tier. The risks are equally clear: the top five customers account for 74.3% of FY25 revenue, the statutory net loss is 21.8 million US dollars, the Global segment is still shrinking, and Oxford Fairwork's rating for crowd pay is only 3/10 (a research rating, not enforcement by Australia's FWO); at a market value of between 2 hundred million and 400 million US dollars it is not in the same league as Scale AI, Surge or Mercor.