India’s AI Boom Has a Human Supply Chain
Automation is reducing routine annotation and increasing demand for specialist evaluation, but India does not track whether workers are moving into higher-skill roles.
News
- Anthropic Tests How Small AI Models Can Replicate Across IoT Devices
- Scientists Call for Global Oversight of Cyber-Capable AI
- Karnataka Eyes Green AI Model Amid Rising Power Needs
- Infosys Sees AI Driving Half Its Work Within Three Years
- FBI Arrests OpenAI Engineer Over Recording Device Planted in Office
- Zoho Freezes Pay Raises as AI Costs Surge
Key Takeaways
01
Estimates of India’s AI data workforce vary widely, from 80,000 to 700,000, because recruiters, platforms and training companies count different groups.
02
Automation is reducing simple labeling while increasing demand for model evaluation, safety testing and specialist knowledge.
03
Companies buying AI data services should require suppliers to disclose worker status, payment terms and career progression to assess workforce stability and labor risks.
Across India, people mark objects in images, transcribe speech, compare chatbot answers and review AI-generated code. Their decisions help companies train and test artificial intelligence. Yet there is no national count of this workforce.
Industry estimates range from 80,000 to 700,000. The spread begins with who gets counted. Some estimates include employees and regular contractors who perform data labeling. Others include people completing occasional online tasks, as well as doctors, lawyers, engineers, and other specialists hired to assess model outputs.
Saravanan Balasundaram, who leads Han Digital Group, a talent and data operations firm, estimates that 80,000 to 100,000 formal and semi-formal workers handle core labeling and annotation. A much larger pool, he says, takes assignments through gig and online platforms.
Neeti Sharma, CEO of TeamLease Digital, a technology staffing company, puts the number of people working directly with AI training data at 150,000 to 200,000. Milind Shah, managing director of digital talent and services firm Randstad Digital India, uses a broader definition and estimates a workforce of 600,000 to 700,000.
This uncertainty matters more as the work changes. Automation is absorbing routine labeling, while more assignments require model evaluation, safety testing and specialist knowledge. India is also investing in domestic AI capacity, and new labor and data-protection rules are bringing worker status and data handling under closer scrutiny. Yet the country does not track whether entry-level annotators move into higher-skill roles. Buyers cannot tell whether their suppliers are building an experienced workforce or relying on a succession of short-term contributors.
Why India’s Workforce Estimates Vary So Widely
Published nearly two years before ChatGPT became publicly available, Nasscom’s February 2021 study captures the business before the surge in generative AI. It valued services delivered from India at about $250 million in fiscal 2020. US clients supplied roughly 60% of revenue, while managed services accounted for 65% to 70% of the market. More than 80% of employees in the workforce models examined came from nonmetro cities.
Nasscom’s most optimistic scenario put the market above $7 billion by 2030, supported by up to one million full-time and part-time contributors.
Ravi Kaklasaria, co-founder and CEO of enterprise training company edForce, cautions against treating any industry estimate as a census. “India does not currently have an official national count of workers involved in AI-training data,” he said.
TeamLease’s estimate has also changed. In July 2025, TeamLease told the Economic Times that the workforce had risen from 20,000 in 2022 to about 70,000. Sharma now puts it at 150,000 to 200,000. TeamLease has not said how much of the difference reflects expansion and how much comes from a broader definition.
Company figures use their own definitions. Social-impact company Karya recruits people from low-income and rural communities for data projects. It said it had reached more than 35,000 people in 2023. Its 2025 report said its Indian workforce had grown from just over 50,000 to 130,000. The totals include people who completed tasks through Karya’s platform and may not have worked continuously.
The same distinction applies to individual projects. Project Vaani collects speech across India’s districts. The Indian Institute of Science and ARTPARK lead the initiative with funding from Google. Karya reports that its part engaged nearly 48,000 contributors in 33 districts. Karya also worked on Project Sanmati, a multi-partner effort supported by the Gates Foundation to expand women’s access to digital work. It says 20,000 women contributed data in six Indic languages across eight states.
A person who completed one batch of tasks may appear in a platform’s workforce figure without holding an ongoing job. Counts rise sharply when they combine salaried annotators, BPO employees, contractors, temporary contributors, online gig workers and outside specialists.
AI Data Work Is Becoming More Specialized
The mix of assignments has changed. Sharma describes an early phase dominated by computer vision labeling, followed by multilingual speech and natural language processing. Current work includes model evaluation, robotics data, and reinforcement learning from human feedback, commonly known as RLHF. In RLHF, people assess model responses, enabling developers to improve how a system behaves.
Shah sees a similar shift in work sent to India. “The scope of work originating from India has fundamentally shifted from early-stage, mechanical tasks like bounding-box tagging toward sophisticated cognitive validation,” he said. The assignments he sees now include debugging code, curating synthetic datasets and refining models for health care, financial services and legal compliance.
Kaklasaria points to prompt-and-response assessment, red-teaming, AI safety testing, and specialist evaluation. These jobs ask workers to find faults, judge context and apply professional knowledge rather than repeat a fixed labeling step.
Recent global skills reports point to stronger demand for advanced capabilities, although neither examines India’s AI data workforce. The International Labor Organization’s August 2026 report found greater demand for cognitive, social, digital and data-science skills. In the World Economic Forum’s Future of Jobs Report 2025, AI and big data led the list of fast-growing skills. Employers surveyed for the report expected 39% of workers’ core skills to change by 2030.
Automation is already absorbing some routine work. Balasundaram says auto-labeling tools and synthetic data can now handle basic two-dimensional boxes and similar tasks. “The human workforce is pivoting toward high-context, complex domain tasks,” he said. India-wide evidence on the pace and employment consequences of this shift is unavailable.
Pay estimates reflect the gap between basic and specialist work, though the sources do not use a common pay measure. Balasundaram puts basic rural annotation at ₹10,000 to ₹15,000, or about $106 to $159, a month. For language evaluation and RLHF, he gives a range of ₹18,000 to ₹30,000 ($190 to $317), and for quality assurance, ₹25,000 to ₹42,000 ($265 to $444). He did not specify whether the latter two ranges refer to monthly salaries or project fees.
Kaklasaria places entry-level annotation at ₹15,000 to ₹30,000 ($159 to $317) a month. Sharma gives ₹22,000 to ₹30,000 ($233 to $317) a month for commodity labeling. She says domain experts on high-value RLHF assignments can earn ₹2,000 to ₹3,000 ($21 to $32) an hour. Shah estimates that niche expertise attracts a premium of 50% to 150%, although he does not identify the comparison wage. The ranges cover different jobs and payment arrangements, so they do not form a single pay scale.
Ankit Aggarwal, founder and CEO of talent engagement and hiring platform Unstop, argues that the idea of a typical data worker no longer fits. “There is no typical worker anymore,” he said. Some people still perform repetitive annotation. Others are doctors, lawyers, engineers or researchers paid for their professional judgment.
“AI is not creating one labor market. It is creating two. One rewards scale and efficiency. The other rewards expertise and judgement.”
— Ankit Aggarwal, founder and CEO of Unstop
Rural Reach and the Missing Career Ladder
India’s annotation business had spread beyond its largest technology centers before generative AI took off. Nasscom recorded that pattern in fiscal 2020. Balasundaram now estimates that 30% to 40% of the basic annotation, localization and domestic BPO work he follows is performed outside tier-one cities. He cites rural BPOs, impact-sourcing organizations and digital cooperatives in Tamil Nadu, Andhra Pradesh, Telangana, Kerala, Uttar Pradesh, Madhya Pradesh, West Bengal and Jharkhand. Kaklasaria also identifies Ranchi, Bhubaneswar, Shillong, Visakhapatnam and Hubballi as emerging locations.
Balasundaram estimates that women make up at least 70% of workers in the structured rural operations he tracks. He links their participation to nearby workplaces, transport and shifts arranged around local needs. A West Bengal experiment found that location influenced whether women accepted a job. The work involved short-term data entry rather than AI annotation.
The May 2026 STEG working paper covered 1,670 households. When the job required office attendance, 15% of women accepted it. When they could work at home while caring for children, 48% accepted. Office-based workers were 3% more accurate and 18% faster. Women who began with paid work at home were more likely to accept less flexible jobs two to three months later. The increase was largest among women who had never worked before. The study shows how flexibility affected participation in that experiment. It cannot establish the same outcome for annotation or model evaluation.
Location may also affect which assignments workers can reach. Shah says tier-two cities can support large teams on structured projects. Bengaluru, he adds, has more people qualified for work requiring greater judgment, including RLHF. No published dataset maps where India’s model-evaluation work is performed.
Foreign buyers remain important. Nasscom attributed roughly 60% of Indian annotation revenue in fiscal 2020 to US clients. Shah puts the overseas share of his team’s work at 75% to 80%. He says more assignments are arriving through global capability centers. “India is the world’s annotation and reinforcement-learning floor, but frontier model-building itself still sits elsewhere in the value chain,” Sharma said.
Kaklasaria declines to put a national percentage on overseas demand because the data is too thin. He also sees more work from Indian companies, startups and public-sector programs. In September 2025, the IndiaAI Mission announced more than 500 data labs and awarded ₹988.6 crore to the IIT Bombay-led BharatGen consortium. The money supports computing infrastructure, models, datasets and skills. Neither announcement specified how many jobs would result, what those jobs would involve or where they would be based.
Contracts Determine Which Protections Apply
Two people doing the same task may have very different legal protections. One may be an employee of a data services company, another a contractor at a BPO, and a third an independent contributor using a platform. The contract and the way the work is organized determine which laws apply.
India’s four Labor Codes took effect on November 21, 2025. The Social Security Code formally recognizes gig and platform workers. Karnataka has gone further in platform work, as covered by its own definitions. Act No. 72 of 2025 was published on September 12, 2025, though took effect on May 30, 2025. Covered platforms must pay a welfare fee equal to 1% to 5% of worker payouts. They must also maintain an internal dispute process and give 14 days’ written notice before termination. The law will not cover every person performing AI data work.
Data protection creates a separate set of responsibilities when projects use identifiable voices, images, health records or financial information. India’s Digital Personal Data Protection Rules, 2025 are being introduced in stages. Rules governing the Data Protection Board and related procedures took effect on November 13, 2025. The framework for registered consent managers begins on November 13, 2026. Most duties that directly affect how companies collect and handle personal data, including notice, consent, security safeguards and breach reporting, begin on May 13, 2027.
That timetable leaves companies a limited period to review their training-data operations. They need to establish how data is collected and which suppliers or subcontractors can access it. The review should identify overseas access and the organization responsible at each stage. Companies must also account for provisions already in force and any other applicable laws.
Shah sees platform annotators outside conventional employment as a major regulatory gap. Aggarwal argues that existing labor categories do not adequately describe people who contribute directly to AI systems. Their concerns are especially relevant when work passes through several contractors and no single company sees the worker’s full experience.
Career movement is equally hard to trace. Balasundaram estimates that 15% to 25% of entry-level rural workers move into quality assurance, project management or specialist evaluation within 18 to 36 months. He estimates that such moves raise earnings by 30% to 60% or more. These figures have not been independently verified. India has no longitudinal study of how many annotators advance, how long progression takes or how many leave after a project.
The limits of visibility are not unique to India. Uma Rani and Morgan Williams analyzed ILO surveys of data workers in India and Kenya conducted in 2022 and 2023. Their peer-reviewed article appeared in the Weizenbaum Journal of the Digital Society in 2026. It documents the extensive human labor behind systems that users often experience as automated.
What Leaders Must Do Differently
C-Suite (CEOs, CIOs and Chief AI Officers) | Large buyers should ask suppliers how many entry-level workers moved into evaluation, safety or quality roles during the previous 12 months. The report should include the associated change in pay and distinguish employees, contractors and platform contributors. By the end of fiscal 2027, management should know whether its spending supports a durable skills base or a succession of short projects. |
Functional Leaders | Procurement and data teams should identify who performs each task, the contractual basis of the work and every subcontractor involved. Payment terms and grievance procedures belong in the same review. Contracts should state how vendors handle qualification tests, rejected tasks and time spent waiting for assignments. Volume alone is a poor measure for work that depends on judgment. Contracts for specialist assignments should also track error rates, training, quality improvement and movement into more demanding roles. |
Boards & Governance | Boards should require a clear account of how each project handles personal data. Management should identify where the data is collected and which suppliers or subcontractors can see it. The account should cover overseas access and name the company responsible at each stage. Legal teams should separate DPDP duties already in force from those beginning in November 2026 and May 2027. They should also determine whether Karnataka’s platform-work law covers any supplier or group of workers. India already performs a broad range of AI data work, from routine labeling to expert review. What the country still cannot show is how often workers move from one end of that range to the other. Suppliers hold much of the information needed to answer that question. Buyers can begin closing the gap by making worker status, pay terms and career movement part of routine vendor reporting. |
Research Context
This article draws on written responses from five industry executives: Saravanan Balasundaram of Han Digital Group; Neeti Sharma of TeamLease Digital; Milind Shah of Randstad Digital India; Ravi Kaklasaria of edForce; and Ankit Aggarwal of Unstop. Their estimates cover workforce size, skills, pay, geography, clients and regulation. The reporting was checked against official records, project disclosures, industry research and an academic field experiment. Where definitions or payment periods differ, the article keeps the figures separate.
