Data annotation services in India: a buyer's guide for AI teams
India is one of the main places AI teams source human data: text and document annotation, image labels, writing ideal answers (SFT), ranking AI answers (RLHF) and expert checks. Choose between crowd platforms, expert networks, large outsourcing firms and managed teams, then run a paid pilot with hidden test items before you scale.
What you can source from India
Human data work has moved up the skill ladder. When Scale AI closed a team of generalist contractors in October 2025, its spokesperson said: "This reflects an industry shift toward higher skill, expert data work." The tasks below range from rule-based labelling to work that needs a graduate's judgement.
| Task | What people do | What it needs |
|---|---|---|
| Text annotation | Classify, tag or extract from text: intent, sentiment, entities, policy categories | Careful reading and consistent rules |
| Document annotation | Pull fields from invoices, forms and records; check them against the source | Attention to detail |
| Image labels | Classify images, draw boxes, check product and document images | A clear guide and visual care |
| Writing ideal answers (SFT) | Write the answer a model should give to each prompt, following a style guide | Strong written English and subject knowledge |
| Ranking AI answers (RLHF) | Compare two or more answers and pick the better one, with reasons | Judgement against a written rubric |
| Expert checks | Graduates in commerce, science or other fields check AI answers in their subject | A degree in the subject and a habit of checking sources |
For the review side in depth, our guide to outsourcing AI answer review and RLHF data has a worked rubric example.
Why AI teams look to India
The market is large and still growing. Grand View Research estimates the global data collection and labelling market at $3.8 billion in 2024, rising to $17.1 billion by 2030. Demand has made some suppliers very large: Reuters reported in July 2025 that Surge AI took in over $1 billion in revenue in 2024.
India already does a share of this work for US clients. A NASSCOM report from February 2021, quoted by the government's IndiaAI portal, put the data annotation market serviced from India at about $250 million in FY2020, with about 60% of revenue from US clients, and said it could exceed $7 billion by 2030.
Time zones help too. India is 9.5 to 10.5 hours ahead of US Eastern time, so a batch sent at the end of your day can be done and checked by your morning.
Four kinds of provider, and when each fits
| Type | How it works | Fits best | Watch for | Examples |
|---|---|---|---|---|
| Crowd platforms | Many independent workers take small tasks; the platform routes and scores | Large volumes of simple, well-defined tasks | Who exactly does your work; quality varies by task | Amazon Mechanical Turk, Appen |
| Expert networks | Vetted specialists, often paid by the hour | Hard reasoning, coding or professional-domain data | High hourly rates; availability by subject | Mercor, Surge AI |
| Large outsourcing firms | Dedicated teams inside a BPO or IT services company | Steady, long-running programmes with fixed processes | Minimum sizes and contract terms | Many Indian BPO and IT firms |
| Managed impact teams | A team hired and trained from under-employed groups, with a quality layer on top | Steady work where you also want a social impact you can report | Ask for evidence on pay and on quality, not just the mission | iMerit, Karya, HerWorkCircle |
Many buyers mix them: a crowd platform for simple volume, and a managed or expert team for the work where judgement matters.
The graduate workforce most buyers miss
Women are about half of India's higher education students: 2.24 crore of 4.50 crore enrolled in 2023-24, according to the government's latest higher education survey. Yet only 28.0% of urban women aged 15 and over were in the labour force in 2023-24, according to the national labour survey. Remote, flexible work is one way to bring more of these graduates into paid work.
This has been done at scale before. iMerit, a data annotation company, has roughly 4,000 staff, a majority of them women. Karya, a non-profit, paid one worker about $5 an hour for Indian-language data work, "nearly 20 times the Indian minimum", TIME reported in 2023.
At HerWorkCircle, every woman takes a skill check that tests reading, writing, attention to detail and judgement of AI answers, and a person on our team reads her answers. For each project, we run a task-specific test before anyone touches live data, and a second person checks the work before it reaches you.
How to measure quality before you scale
- Hidden test items. Mix items with answers you already know into each batch. They show accuracy without anyone knowing which items are scored.
- Agreement. Have two people do the same items separately. A widely cited guide to the kappa statistic says any kappa below 0.60 means inadequate agreement; low agreement usually points to unclear guidelines before it points to the people.
- Agreement per task, not averaged. Ask any vendor for agreement broken down by label or task type, so a weak area can't hide inside a good average.
- Your own review time. Count the hours your team spends checking. A cheap batch that takes a day to fix is not cheap.
Plan your pilot here. The shares and pass marks are starting points that you should adjust for your task.
A pilot of 500 items: classifying, tagging or extracting from text.
- Practice round
- 25 items, all checked by you, before the pilot starts
- Hidden test items
- 50 with answers you already know, mixed into the batch
- Done twice
- 100 by two people separately, to measure agreement
- At least 95% of the hidden test items are right.
- Agreement on the double-done items is at least what two of your own people reach on the same items.
- Days 1 to 2 Send the guidelines and examples. The team does the practice round; you answer every question and fix the guidelines.
- Days 3 to 7 The pilot runs. Check a sample each day and send feedback the same day.
- Days 8 to 10 Score the hidden items and the agreement, count your own review hours, and decide.
- Guidelines with a definition and an example for every label
- 20 worked examples, including hard and borderline ones
- Your hidden test items, with answers
Security questions to ask any vendor
- Who exactly will see the data, and have they signed a confidentiality agreement?
- Does the work happen in your tools and accounts, or will data be copied to theirs?
- Can any part of the work be passed to another company or to freelancers without your written consent?
- What security certifications do they hold, and what do those certifications actually cover?
- How is your data deleted at the end, and will they confirm it in writing?
- Can personal data be removed or masked before it reaches the team?
Our guide to the risks of outsourcing to India explains how India's data protection law treats work done for foreign clients, and who owns the work you pay for.
Ethics: what to ask, with evidence
In January 2023, TIME reported that OpenAI's work to make ChatGPT less toxic used outsourced Kenyan workers earning less than $2 an hour. Since then, buyers have been asked harder questions about their data supply chain.
The Partnership on AI, an industry group, published guidelines for responsible sourcing of this work, developed with DeepMind and adopted by it across the organisation. In short:
- Pay workers above the local living wage.
- Design and run a pilot before launching a project.
- Choose workers suited to the task.
- Give verified instructions or training materials.
- Set clear, regular ways to communicate with the workers.
Ask any vendor what the person doing your task is paid, when they are paid, and whether rejected work is paid. A vendor that is proud of its answers will give them.
How human data work is priced
| Model | Fits best | Ask about |
|---|---|---|
| Per item or label | Well-defined labels with a steady time per item | What happens to items that fail checks |
| Per hour | Writing, ranking and expert work, where time per item varies | How hours are tracked and capped |
| Monthly team | A steady programme with a fixed team that learns your data | Cover for leave, and how fast the team can grow |
Get the price after the pilot, not before: the pilot tells both sides how long each item really takes.
How a pilot with HerWorkCircle works
- Send a brief: the task, a few examples and the volume you expect. Our brief builder helps.
- On a free 20-minute call, we agree the task, the pass rules and a clear price for the pilot.
- We pick women whose skill check fits the task and give them a task-specific test before any live data.
- The team does a practice round and we bring your answers to their questions back into the guidelines.
- The pilot runs, and a second person checks every piece before it reaches you. You score the hidden test items and decide.
We take on English text work, document and image labels, and subject checks by graduates. Send a brief to start.
Questions
Which companies provide data annotation services in India?
Four kinds: crowd platforms such as Appen, expert networks such as Mercor and Surge AI, large Indian BPO and IT firms, and managed impact teams such as iMerit, Karya and HerWorkCircle. Choose by task difficulty and volume.
How much does data annotation cost in India?
It depends on the task and the pricing model: per item, per hour or a monthly team. Run a small paid pilot first, because it shows how long each item really takes.
What is SFT data?
Supervised fine-tuning data: prompts paired with the ideal answer, written by people, used to teach a model how to respond.
How do you measure data annotation quality?
Mix hidden test items with known answers into each batch, have two people do some of the same items to measure agreement, and track your own review time.
Can graduates in India do expert data work for AI?
Yes, for tasks that match their subject and English level. Test them on your task first, and check a sample of their work during a pilot.
How do I know a data vendor treats its workers fairly?
Ask what the person doing your task is paid and when, whether rejected work is paid, and compare the answers with the Partnership on AI's responsible sourcing guidelines.
Sources
- Scale AI cuts more contractors (16 October 2025) Business Insider, via AOL
- Data collection and labeling market size report Grand View Research (estimate)
- Exclusive: Scale AI's bigger rival Surge AI (1 July 2025) Reuters, via AOL
- Data annotation in India heightening with the AI revolution (NASSCOM report figures) IndiaAI, Government of India
- AISHE 2023-24: higher education enrolment reaches 4.50 crore (8 July 2026) Careers360
- Periodic Labour Force Survey annual report 2023-24: press note (22 September 2024) Ministry of Statistics and Programme Implementation
- iMerit: from Silicon Valley to Salt Lake (31 March 2024) The Quint
- The workers behind AI rarely see its rewards (Karya, July 2023) TIME, via Yahoo
- Interrater reliability: the kappa statistic (2012) McHugh, Biochemia Medica, via PubMed Central
- OpenAI used Kenyan workers on less than $2 per hour (18 January 2023) TIME
- Responsible sourcing of data enrichment services Partnership on AI