Why Appen can’t keep up with Mercor, TrueUp, and Scale AI
Appen is struggling to keep up with newer data-labeling players like Mercor, TrueUp.io, and Scale AI for deep structural reasons, while the entire sector is suddenly booming because the generative-AI wave created a brand-new, high-margin “expert-in-the-loop” market that barely existed three years ago.
🧱 1. What Appen Was Built For vs. What AI Labs Now Need
Appen’s 25-year-old model is essentially “crowd + project manager.”
It keeps a million-person global click-worker pool, parcels out micro-tasks, and manages it through a web UI.
That setup works great for classic supervised learning — drawing boxes around cars or transcribing short audio clips — tasks priced at a few cents per label.
But GPT-era training needs something else entirely:
PhD-level raters who can spot subtle hallucinations, judge constitutional alignment, write chain-of-thought explanations, or verify multi-step agent outputs.
These jobs pay $40 – $120 / hour, not 4¢. Appen’s crowd pool and cost structure were never built for that tier, so their quality scores on LLM evaluation work have lagged behind boutique rivals.
⚙️ 2. Technology Stack Mismatch
The new generation, Scale, Mercor, TrueUp, are API-first data engines.
They offer REST / GraphQL endpoints, Python SDKs, CI / CD hooks, and model-in-the-loop pre-labels. You can plug them into a training pipeline in ten lines of code.
Appen remains UI-first. Clients upload spreadsheets, and project managers route tasks manually. Integrating Appen into an MLOps loop means weeks of setup and a services contract, exactly what modern AI teams try to avoid.
⚡ 3. Speed and Automation
Scale and the newcomers pre-label with their own models, letting humans fix only the edge cases. Turnaround times are measured in hours.
Appen, even with its AI-assist add-ons, still depends on full human passes. Projects that Scale finishes in 24 h often take Appen two weeks.
🧩 4. Enterprise Perception & Governance Headaches
After Meta acquired Scale’s parent entity, big-tech teams began dual-sourcing to avoid lock-in. Appen could have been the logical second source, but many failed security audits because its annotation happens on unmanaged home laptops.
Former managers say quality-control shortcuts, like raters running auto-scripts to hit quotas, cost Appen its $82 M Google contract and halved its enterprise client base within 18 months. Once that trust eroded, switching became easy.
💸 5. Economics That No Longer Scale
Appen’s gross margins are crushed by contributor support and heavy PM overhead.
The new entrants run lean, software-style ops: small internal teams, on-demand expert pools, and automation. They can pay raters more while still undercutting Appen on price.
🚀 6. Why the Startup Wave Is Happening Now
- 🤖 ChatGPT’s launch triggered a surge in RLHF / red-team budgets — OpenAI alone has spent nine figures on human-rater work since 2023.
- 🧠 Every model builder (Anthropic, Cohere, Mistral, Adept, etc.) now needs tens of thousands of expert hours but doesn’t want to build in-house rater orgs.
- 💼 That opened a $4–5 B niche for high-skill data-labeling marketplaces, exactly what Mercor, TrueUp, Surge AI, Outlier.ai, and Micro1 were founded to serve.
- 🏠 Remote work normalization means credentialed experts (doctors, lawyers, STEM grads) are happy to log in part-time at $75 / hr, unthinkable pre-COVID.
- 💰 Venture money is flooding in because these firms show software-like gross margins (40–60 %) versus Appen’s 15–20 %.
🔚 Bottom Line
Appen’s legacy architecture, low-skill workforce brand, manual PM layer, and quality scandals make it ill-suited for the expert annotation gold rush that generative AI unleashed.
Scale AI saw the shift early and rebuilt around API + model-in-the-loop.
Mercor, TrueUp, and the 2023–24 wave of startups were born for this new expert-rater marketplace, and they’re capturing the fastest-growing, highest-margin slice of the data-labeling budget.
📉 Meanwhile, Appen’s revenue keeps sliding.
