Every enterprise AI system that works in production rests on a layer most boardroom conversations skip: people deciding what a piece of data means. Someone draws the box around a corroded pipe joint, tags the clause in a loan agreement, transcribes the Bhojpuri phrase a caller used, or ranks two model answers for a legal research tool. That work, usually called data annotation or labelling, is where model quality is set long before anyone measures it.
Data annotation in India is now a sizeable export industry and a fast-growing domestic one. It is also a labour market with real questions about pay, security and recognition. For the CIO or chief data officer buying annotation, the decisions look technical (accuracy targets, tooling, turnaround), but they are as much about operating models and supplier ethics.
This piece does three things. It sets out what the credible public numbers say about the industry and what they do not. It offers a simple quality, cost and speed model, with every assumption stated, that buyers can use to price annotation honestly. And it gives a factual account of labour conditions, because a supply chain that cannot survive scrutiny is a reputational risk, not a cost saving.
What the numbers on data annotation in India actually say
The most cited figures come from Nasscom's Data Annotation Report of February 2021, summarised by the government's IndiaAI portal in January 2022. Nasscom valued India's data annotation market at about USD 250 million in FY20, with roughly 60% of revenue coming from US clients. Managed services, where a vendor runs the workforce, tooling and quality control on the client's behalf, made up 65% to 75% of the market. The report said the market "can exceed up to USD 7 Bn by 2030."
Two cautions apply. First, the FY20 base predates the generative AI wave. Demand since 2022 has shifted from image and text tagging towards preference ranking, red-teaming and expert review for large language models, and no public Nasscom update of comparable rigour has replaced the 2021 baseline. Second, "up to USD 7 Bn" is a potential, not a forecast with a central estimate. It is useful as a sense of direction, not as a planning number.
The workforce numbers are thinner still. A Thomson Reuters Foundation report published in March 2023 cited a Nasscom projection of up to one million full and part-time data annotation workers in India by 2030. The same report said about 50,000 annotators work as freelancers on platforms such as Amazon Mechanical Turk and Clickworker, and about 20,000 work full-time for third-party firms. Those figures are a snapshot from one reported source and should be treated as orders of magnitude.
Industry maturity is the more telling indicator. A 2021 analysis hosted on Nasscom's community platform estimated that about 75% of Indian annotation players were in initial or growth phases, while about 70% of global players were in intermediate to advanced phases. In plain terms, India had scale in labour and cost, but fewer firms with mature quality systems, tooling and domain specialisation.
The work has changed: from bounding boxes to expert judgement
The early annotation economy was built on tasks that any trained worker could do with modest English and computer skills: drawing boxes on images, transcribing audio, tagging sentiment. That work still exists and still matters for computer vision in factories, retail and road inspection.
Three shifts have changed the mix.
Expert annotation. Models that support clinicians, lawyers, auditors or maintenance engineers need labels from people who understand the domain. A radiology label from a general annotator is a different product from one checked by a radiologist. Hourly costs rise by an order of magnitude, and throughput falls.
Preference and evaluation data. Training and tuning large language models requires humans to compare outputs, rate them against rubrics and write ideal answers. This is closer to editorial work than to labelling, and its quality depends on rubric design as much as on the annotator.
Indian-language and voice data. Enterprises building for Indian customers need speech and text in many languages, dialects and code-mixed forms. Collecting and labelling that data requires annotators who are native speakers, and quality control that understands regional variation.
The common thread is that the cost of a label is increasingly driven by the cost of judgement, not the cost of a click. That changes how annotation should be priced and bought.
A quality, cost and speed model for annotation
Buyers often compare annotation vendors on a headline price per label and a headline accuracy figure. Both numbers hide the variable that matters most: how much review the vendor does, and how good that review is. The model below makes that explicit. It is illustrative. The parameters are assumptions chosen to show the shape of the trade-off, not observed market rates.
The model. For each label:
Base cost = loaded hourly cost of an annotator (C) divided by labels completed per hour (L). A share of labels, the review coverage (r), is checked by a reviewer. Review takes a fraction (m) of the original effort. Each first-pass error (rate e) that a reviewer catches (catch rate k) is redone at the cost of one more label. Cost per label = (C ÷ L) × (1 + r × m + e × r × k). Residual error after review = e × (1 − r × k).
Assumptions for four operating points (illustrative).
Outputs.
Three readings follow.
First, quality is mostly bought through review, not through the first pass. Moving from A to C roughly triples the cost per label but cuts residual error from about 9% to 0.5%. A vendor quoting a low price and high accuracy at the same time should be asked what review coverage and catch rate sit behind the number, and how accuracy was measured.
Second, expert work changes the unit economics entirely. At operating point D, a label costs more than twenty times what it costs at B. For most enterprise problems the sensible design is a hybrid: generalist annotators do the volume, and experts adjudicate the hard or ambiguous cases. The share routed to experts becomes the main cost lever.
Third, speed is set by the scarcest skill. A crowd can be scaled in days. A pool of qualified reviewers or domain experts cannot. Deadlines for annotation projects are usually missed at the review stage, not at the first pass.
The model also gives buyers a way to decide how much quality to pay for. If a mislabelled item costs the business a known amount downstream, for example a false negative in a defect detection system or a misrouted customer complaint, then the value of reducing error by one percentage point across a dataset can be set against the extra cost of review. Where that downstream cost is small, operating point B may be rational. Where it is large, C or a hybrid with D is cheaper in total. A separate PanScience Research piece covers how to evaluate models before deployment, which is where label quality shows up as model quality.
Labour conditions: what the evidence shows
Annotation is often described as part of India's gig and platform economy, and the scale of that economy is well documented. NITI Aayog's June 2022 report, India's Booming Gig and Platform Economy, estimated 7.7 million gig workers in 2020-21, about 1.5% of the workforce, and projected 23.5 million by 2029-30, or 4.1% of the workforce. It classified about 47% of gig jobs as medium- skilled, about 22% as high-skilled and about 31% as low-skilled. Annotation sits across all three bands depending on the task.
Globally, a World Bank report published in 2023, Working Without Borders: The Promise and Peril of Online Gig Work, estimated between 154 million and 435 million online gig workers, or 4.4% to 12.5% of the global labour force. The upper estimate includes people for whom gig work is a secondary or marginal activity. The Thomson Reuters Foundation report cited earlier noted an International Labour Organization estimate that India accounts for about a third of the world's online freelance workers.
On pay and conditions, the most specific public reporting is journalistic rather than statistical. The same March 2023 report described an engineering graduate in Salem working on Mechanical Turk who earned between USD 10 and USD 30 a day, with only about two in ten of his tasks approved. A founder of an annotation firm told the reporters that firms typically paid fixed monthly salaries of about USD 200 to USD 300, and described ₹15,000 to ₹25,000 a month as a very good salary in a village. A worker employed by a Kerala firm after two months of training reported about USD 220 a month on a regular day shift. The report also cited a 2021 IT for Change study which found that fewer than a quarter of Indian workers on Mechanical Turk were women, and that women often ended up with the lowest-paying, most tedious tasks.
Three features of the work stand out from this evidence. Platform work is paid by accepted task, so rejection rates directly cut earnings and workers bear the risk of unclear instructions. Employment through annotation firms offers more stable pay, often in smaller towns where it compares well with local alternatives. And content that is distressing, such as violent or abusive material used to train safety filters, carries psychological risk; the Thomson Reuters Foundation report cited a Time investigation into such work being outsourced to workers in Kenya at under USD 2 an hour.
The regulatory position changed in late 2025. The four labour codes, including the Code on Social Security, 2020, took effect from 21 November 2025, according to EY's alert on their implementation. The Code allows central and state governments to frame welfare schemes for gig and platform workers covering life and disability, health and maternity, and old-age protection, and requires aggregators to contribute 1% to 2% of annual turnover, capped at 5% of amounts paid or payable to gig workers. EY noted that several supporting central and state rules were still to be notified. How far these provisions reach annotation workers will depend on those rules and on whether a given arrangement is treated as platform work, employment or a services contract.
None of this makes annotation bad work. For many workers it is a route into the digital economy that did not exist a decade ago. But it does mean that the buyer of annotation is buying a labour practice along with a dataset.
What buyers should ask for
The questions below translate the model and the labour evidence into procurement terms. They are designed to be asked of any vendor, in India or elsewhere.
How the practice is evolving
The most useful change in the market is the move from annotation as a commodity task to annotation as part of a data operations function that sits alongside model operations. In this model, annotation guidelines, gold sets, review workflows and model error analysis are owned together, so that what the model gets wrong in production changes what annotators are asked to do next.
PanScience's portfolio includes two ventures that work in this layer, and they illustrate the split. Indika AI describes itself as a data-centric AI and model operations business and reports annotation accuracy of about 98% (company-reported). FlexiBench describes itself as a managed AI workforce with more than 30,000 annotation experts (company-reported). Read against the model above, the useful questions for any such provider, including these, are how accuracy is defined and audited, what share of work is reviewed, and how the expert pool is qualified and paid.
For enterprises, the choice is less between in-house and outsourced annotation than between owning the data operations function and renting it. Organisations whose models are core to their products tend to keep guideline ownership, gold sets and acceptance testing in-house, even when the labelling itself is outsourced. Those that treat annotation as a one-off project often discover, when the model degrades, that nobody owns the definition of a correct label.
Conclusion
India has the workforce, language range and cost base to remain a leading supplier of annotation, and a growing domestic market as Indian enterprises build their own models and applications. The public data on the industry is older and thinner than its importance warrants, and anyone planning around a single market figure should treat it with care.
For buyers, three points matter more than the headline price. Quality is bought through review, and review must be specified. Expert work changes the economics and should be routed deliberately, not by default. And the annotation supply chain is a labour supply chain, now within reach of the Code on Social Security, which deserves the same scrutiny as any other part of an enterprise's sourcing.
Methodology and sources
Market figures come from Nasscom's Data Annotation Report (February 2021) as summarised by the IndiaAI portal (indiaai.gov.in/article/data-annotation-in-india-heightening-with-the-ai-revolution, published 13 January 2022) and from a 2021 analysis on Nasscom's community platform (community.nasscom.in/communities/ai/data-annotation-and-labelling-services-india-value- proposition-and-industry-maturity). Workforce estimates, pay reports and the ILO and IT for Change references come from the Thomson Reuters Foundation's Context report "AI boom is dream and nightmare for workers in Global South" (context.news/ai/ai-boom-is-dream-and-nightmare-for- workers-in-global-south, March 2023); these are reported figures and individual accounts, not survey statistics. Gig economy figures come from NITI Aayog's June 2022 report as reported by the government's IndBiz portal (indbiz.gov.in/indias-gig-economy-to-have-23-5-million-workers-by-2029- 30-niti-aayog). The global online gig work estimate is from the World Bank's 2023 report Working Without Borders, as reported by Staffing Industry Analysts (staffingindustry.com/news/global-daily- news/world-bank-estimates-there-are-435-million-online-gig-workers-globally). Labour code provisions come from EY India's November 2025 alert (ey.com/content/dam/ey-unified-site/ey- com/en-in/alerts-hub/2025/11/new-labour-codes-implemented-across-the-country-effective-21- november-2025.pdf). The quality, cost and speed model and all its parameter values are illustrative assumptions by PanScience Research, not observed rates. PanScience venture descriptions are company-reported from panscience.xyz.