Appen built one of the world's largest human-powered data annotation platforms from Sydney, supplying the labelled training data that major AI models have relied on for years. At its 2021 peak, the ASX-listed company commanded a market capitalisation of around $4 billion and counted Google, Meta, and Microsoft among its largest customers. By 2024 those same customers had pulled back sharply. Appen's revenue dropped from over AU$600 million in 2021 to below AU$300 million two years later, and the share price followed. The question now is whether Appen can reinvent itself fast enough to survive the very AI boom it helped create.
What Appen actually does
Appen's core business is human-in-the-loop data services: recruiting, managing, and paying a global crowd of annotators to label images, transcribe audio, evaluate search results, and rank AI-generated outputs for quality. That work is unglamorous but essential. A language model needs millions of labelled examples to learn the difference between sarcasm and sincerity, between a safe answer and a harmful one. Appen's crowd of over one million workers across 170 countries has historically done that job at scale.
The company also operates a relevance and search evaluation business, where human raters assess whether a search engine's results match a user's intent. Google has been the single largest customer here for most of Appen's listed life, at times representing more than 60 percent of total revenue. That concentration was always a structural risk. When Google started investing more heavily in its own internal labelling operations and synthetic data pipelines, Appen felt it immediately.
Why the revenue collapsed so quickly
Three forces converged to undermine Appen's position. First, its hyperscaler customers cut external data annotation spend sharply through 2022 and 2023 as their own AI investments matured and their internal capabilities grew. Second, the rise of synthetic data generation offered a cheaper and faster alternative for certain labelling tasks. Rather than paying human workers to label 100,000 images, a model can generate and auto-label synthetic variants in minutes. Third, the rapid improvement of foundation models reduced the volume of hand-labelled data required for fine-tuning, at least for well-resourced labs with large pre-training corpora.
Appen has been candid in its ASX filings about the customer concentration problem. It's worth noting that the company's five largest customers have consistently accounted for more than 80 percent of revenue, making any single contract loss a material event. When Google and Meta both reduced spend in the same financial year, Appen had little buffer. The broader pattern here echoes what has happened to marketplace businesses like Airtasker: building on top of a few large demand-side relationships creates efficiency but leaves the platform dangerously exposed when those relationships shift.
The pivot: from crowd labour to AI-powered annotation
Appen's response has been a pivot it calls the "New Appen" strategy. The core idea is to move up the value chain. Instead of competing primarily on cost-per-annotation by deploying cheaper crowd labour, Appen is investing in its own AI-assisted annotation tools that accelerate human labellers. A human annotator using AI pre-labelling can process images or documents far faster, reducing the effective cost while maintaining quality signals that fully automated pipelines struggle to match.
Appen has also expanded into generative AI data services specifically, helping clients build datasets for RLHF (reinforcement learning from human feedback) and evaluating the safety and factual accuracy of large language model outputs. This is a genuine growth area. As enterprises deploy AI into regulated contexts, the demand for careful, human-reviewed evaluation data is rising. Appen's crowd infrastructure, with its language diversity and domain specialisation, is genuinely hard to replicate quickly.
The company has been direct about targeting mid-market enterprise clients rather than relying on hyperscaler budgets. It has also built partnerships with AI platform vendors to embed its services into annotation workflows rather than selling discrete labelling projects. Whether these moves generate enough new revenue fast enough is the central question.
Synthetic data: threat and opportunity
Synthetic data is both Appen's biggest competitive threat and a potential revenue stream. The threat is obvious: if a model can auto-generate labelled training data, the need for human annotators shrinks. The opportunity is subtler. Synthetic data needs validation. A synthetic dataset that embeds biases or errors into a production model causes real harm. Human review of synthetic data outputs, done at scale and with domain expertise, is a service Appen is positioned to offer.
This dynamic is worth watching because it mirrors broader debates happening inside enterprise AI teams right now. The question of when to trust automated data pipelines and when to insist on human oversight is one that Australian enterprises deploying AI are grappling with directly. It connects to the same tensions explored in discussions around when synthetic data helps and when it backfires: the cost savings are real, but the failure modes are subtle and sometimes catastrophic.
The balance sheet reality
Appen entered 2025 with a weakened balance sheet and a reduced workforce. The company shed hundreds of staff across its corporate operations and made changes to its executive team. It raised capital through a placement in 2024 to shore up liquidity while the restructuring took hold. Those moves bought time, but the company's path back to the revenue scale it once held is not straightforward.
One realistic scenario is a strategic acquisition. Appen's global crowd infrastructure, its annotation tooling, and its enterprise customer relationships have genuine value even in a compressed market. Several private equity and strategic buyers have been reported as taking interest in ASX tech companies trading at distressed multiples. Appen fits that profile. The ASX tech sector in 2026 has seen consolidation pressure build across companies that grew fast on hyperscaler dependency and are now being repriced. Appen sits squarely in that group.
What comes next
The honest answer is that Appen's future depends on whether its New Appen pivot generates new customer commitments before its cash position forces further dilution or a sale. The data services market for generative AI is growing. The global demand for RLHF-quality human feedback, multilingual training data, and AI output evaluation is not disappearing. Appen has the operational infrastructure to serve that demand. What it lacks is the customer diversification it needed years ago.
For Australian IT and investment observers, Appen is a clear case study in the risks of building a services business on top of a concentrated set of hyperscaler relationships. The company rode that concentration to extraordinary growth and is now working hard to survive the pullback. Whether it threads the needle will say a great deal about the durability of the human-in-the-loop model in an era of rapidly improving AI automation.