Data labeling sits between raw information and usable artificial intelligence. The available figures show a global service ecosystem with contributor networks exceeding one million people, support for hundreds of languages, and a measurable economic footprint in the United States. They also show persistent pressure around accuracy, diversity, sourcing, and the speed of model updates.
Contents
- The scale of data labeling providers
- Languages, countries, and data types
- Accuracy, diversity, and annotation bottlenecks
- The US economic footprint
- Who participates in labeling work
- Business and delivery capacity
The scale of data labeling providers
Several provider-reported figures illustrate the size of the data-labeling and AI-training-data ecosystem. Appen says it has more than 30 years of experience in AI training data, more than 1 million vetted contributors worldwide, and representation in more than 170 countries. Its “Appen about us” figures also report support for more than 235 languages and more than 20,000 completed projects. Appen says 80% of leading large language model builders are customers.
TELUS Digital reports a similarly large contributor network. Its Fine-Tune Studio release says the TELUS Digital AI Community has more than 1 million crowdsourced contributors across 150 countries. The release says Fine-Tune Studio creates datasets in more than 100 languages and supports four data types: text, audio, image, and video.
These figures describe provider-reported reach rather than a single audited global market total. They are useful for showing operating scale, geographic coverage, and the range of work that labeling platforms say they can coordinate.
The provider landscape also has a formal competitive benchmark. TELUS Digital says Everest Group evaluated 19 providers in its 2024 PEAK Matrix assessment, with only five recognized as Leaders. That result places data services in a market where geographic reach and contributor volume coexist with substantial differences in evaluated provider position.
Languages, countries, and data types
Language coverage is a central measure of labeling capacity because models used across regions need training and evaluation material that reflects more than one dominant language. Appen reports more than 235 supported languages and contributors represented in more than 170 countries. TELUS Digital reports more than 100 languages for Fine-Tune Studio and a contributor community spanning 150 countries.
The two provider descriptions are not directly interchangeable. Appen’s figures refer to its broader AI-training-data operation, while TELUS Digital’s language figure is attached specifically to Fine-Tune Studio datasets. Still, together they show why labeling is often organized as a distributed collaboration problem: the work may require language specialists, regional knowledge, and reviewers across many locations.
Fine-Tune Studio also lists four data types: text, audio, image, and video. The platform release lists four fine-tuning or alignment approaches as well: supervised fine-tuning (SFT), reinforcement learning from human feedback (RLHF), direct preference optimization (DPO), and red teaming. These are categories of supported work, not a measurement of how frequently each approach is used.
Innodata’s Q2 2024 investor presentation gives another view of operational reach. The company said it had more than 20 delivery locations and more than 5,000 global experts. It also described its Goldengate platform as state-of-the-art for more than 50 knowledge tasks. These figures suggest that labeling-related production can extend into specialized and domain-specific data work rather than remaining limited to simple image or text categorization.
Accuracy, diversity, and annotation bottlenecks
Appen’s 2024 State of AI report surveyed more than 500 IT decision-makers. In that survey, bottlenecks related to sourcing, cleaning, and annotating data rose 10%, while reported data accuracy fell 9% and data availability challenges increased 7%. The figures point to a difficult operating environment: organizations may have more demand for AI systems while experiencing more friction in preparing the data those systems require.
The same report found that 97% of respondents agreed data diversity, bias reduction, and scalability are vital. It also found that 80% emphasized the importance of human-in-the-loop processes. In this context, human review is not merely a staffing detail. It is presented as part of the quality and governance process for data used to build or refine AI systems.
Earlier Appen research shows that these concerns were already visible in 2022. Appen’s 2022 State of AI report was based on 504 interviews. It found that 51% of participants said data accuracy was critical to their AI use case, while 80% said data diversity was extremely important or very important. The report also found that 95% agreed synthetic data would be a key player for inclusive datasets.
The 2022 report identified sourcing as a practical obstacle: 42% of technologists said data sourcing in the AI lifecycle was very challenging. It also found that 90% were retraining their models more than quarterly. Frequent retraining increases the importance of repeatable data pipelines, because labels and evaluation material may need to be refreshed as models, products, and requirements change.
| Measure | Reported result | Source and period |
|---|---|---|
| Respondents calling diversity, bias reduction, and scalability vital | 97% | Appen, 2024 State of AI report |
| Respondents emphasizing human-in-the-loop processes | 80% | Appen, 2024 State of AI report |
| Participants saying accuracy was critical to their use case | 51% | Appen, 2022 State of AI report |
| Respondents rating data diversity extremely or very important | 80% | Appen, 2022 State of AI report |
| Respondents agreeing synthetic data would support inclusive datasets | 95% | Appen, 2022 State of AI report |
| Technologists calling data sourcing very challenging | 42% | Appen, 2022 State of AI report |
Appen’s 2020 State of AI report provides an older reference point, and its figures should be read as historical findings rather than current measurements. That report said 70% of companies reported C-suite involvement in AI projects, nearly 75% considered AI critical to success, and almost 50% felt their company was behind on its AI journey. It also said executive visibility and involvement in AI had risen by more than 30% year over year.
The same 2020 report said the percentage of companies investing more than $5 million in AI had effectively doubled. Two-thirds of companies did not expect any negative impact from COVID-19 on their AI strategies, while nearly 50% accelerated those strategies and 20% accelerated them significantly. These results describe the business environment reported at that time, not a forecast of present-day investment.
The US economic footprint
Scale’s economic impact study estimates that the US data annotation industry contributed $5.7 billion to US gross domestic product in 2024. The study projects total US industry impact will reach $19.2 billion by 2030. The 2030 figure is a projection, so it should not be treated as a measured 2030 outcome.
Scale says data annotation work supported nearly 200,000 flexible earning opportunities in the United States in 2024. It also says the work sustained about 9,000 full-time employee jobs directly and supported another 25,000 full-time employee jobs through suppliers and local spending. Those categories describe different parts of the reported employment impact; they should not be collapsed into one claim about a single headcount.
The study also reports more than $1.2 billion in federal, state, and local tax revenue generated by the industry in 2024. Taken together, the GDP, employment, and tax figures frame annotation as an economic activity with effects beyond the organizations commissioning labels. The figures remain estimates from Scale’s economic impact study, rather than a government census of every labeling transaction.
Who participates in labeling work
Scale’s economic impact study reports several characteristics of respondents. It says 84% held at least a bachelor’s degree, about 65% were under 44, and 94% were also engaged in other work, education, or caregiving responsibilities. It also says 2 in 10 contributors reported living with a disability.
These figures describe the respondent population reported by Scale and should not be generalized to every person performing data-labeling work. They do, however, highlight the variety of arrangements that can exist around flexible annotation opportunities. A contributor may combine labeling with employment, study, or caregiving rather than treating it as a standalone occupation.
The reported disability figure is especially relevant to discussions of access and flexibility, but it is still a respondent statistic, not a universal characteristic of the industry. Likewise, the education and age figures indicate the composition measured in the study without establishing a causal relationship between those characteristics and labeling quality.
Business and delivery capacity
Public-company results provide additional indicators of the organizations operating around data services and AI support. TaskUs reported $227.5 million in total revenue for Q1 2024, $11.7 million in GAAP net income, a 5.1% GAAP net income margin, and $50.6 million in adjusted EBITDA. Its adjusted EBITDA margin was 22.2%. TaskUs reported 49,600 teammates at the end of Q1 2024 and operations across 27 locations in 12 countries as of March 31, 2024. Its 2024 revenue outlook ranged from $925 million to $950 million.
Innodata reported $26.5 million in Q1 2024 revenue and 41% year-over-year revenue growth. It raised its 2024 revenue guidance to at least 40% organic year-over-year growth. In its Q2 2024 investor presentation, Innodata said it had contracts with five of the Magnificent 7 and two additional Big Tech companies year to date in 2024. The presentation also said its Media Intelligence Platform served about 1,500 customers and its Medical Data Intelligence Platform served about 12 customers.
Innodata’s investor deck described 85 plus 6 adopters across enterprise and large-tech categories. Because the presentation uses that “85+6” formulation, it is best retained as reported rather than converted into an unsupported combined interpretation.
Historical Appen findings add context on operating cadence. Its 2020 report said three out of four companies updated models at least quarterly, and 40% of quarterly updaters felt that lack of data or data management was a challenge. Appen’s 2022 report similarly said 90% were retraining models more than quarterly. These are findings from separate reports and periods, but both identify data operations as a recurring part of AI development.
Appen’s 2022 report also found that 49% of business leaders said their organization was ahead of others in its industry, while another 49% said it was even with others. Those responses leave no quantified remainder to describe as behind, so they should be presented only as the reported “ahead” and “even” shares.
The combined picture is a data-labeling field built around large distributed communities, multilingual and multimodal production, and continuing pressure to improve accuracy and inclusion. Its reported US economic contribution, provider capacity, and survey results show both the scale of the work and the operational problems that labeling teams are expected to solve.