Data annotation jobs are becoming increasingly vital in providing jobs to people from tier 2 and 3 cities in India.
Data annotation jobs are where workers repeatedly feed data into AI models and direct them to understand it correctly.
India’s data annotation tools market was worth $138 million in 2026 and is expected to reach $648 million in 2031, according to Ken Research.
While data annotation is currently providing job opportunities to a lot of people, there are fears that as models become more sophisticated, it will remove the need for the current workers in the field.
In a conversation with Invezz, Radha Ramaswami Basu, Executive Vice President and Head of iMerit, an EXL Company, discusses how the growing adoption of artificial intelligence is changing the nature of data annotation jobs rather than simply eliminating them.
She explains how workers in India’s tier 2 and tier 3 cities can move from routine labelling tasks to specialised roles in model evaluation, red teaming and human oversight, while examining the opportunities and challenges facing India’s AI workforce.
Invezz: There are many fears that AI is going to replace jobs, and that includes people who annotate data as models become more sophisticated and perfect. There are also fears that AI increases productivity, and fewer people are required to do a task that needed more people in the past. How would you respond to these claims? Could you provide a counterargument to this claim?
Realistically, no one has been able to state for sure how this will play out.
We can provide some observations from being a part of the industry with a focus on human experts and their role in the AI stack.
First, AI is heavily dependent on human input, human expertise, and human oversight.
We do not expect this to go away. But we expect that this will become more advanced, more nuanced and require both judgment and nuance.
In our own space, routine data labeling, such as drawing boxes around cars, is being automated.
People who previously did that work will need to move to harder tasks where the situations are more complex and multidimensional.
Instead of assessing whether something is a car in an image, the expert has to assess what actions the car may and should take or what to do if there is a sudden emergency.
The work becomes more descriptive and subjective, and checking the work becomes even more advanced. Knowing what “right answer” looks like becomes non-trivial.
As a model becomes more capable, the cost of its mistakes rises, and so does the need to check it.
A model that drafts emails needs light review. A model that approves a loan, reads a scan, or steers a car needs people who can show it was right.
Every model update also has to be evaluated again, because a fix in one area can break another.
Demand is moving toward expertise. A few years ago, most requests we received were for large teams to label images.
Today, clients increasingly ask for physicians to check a model’s clinical reasoning, lawyers to test its reading of a contract, and mathematicians to find the flaw in a proof.
Coders are writing fewer basic functions and instead are reviewing, correcting and approving more complex behaviours generated by AI.
A labor market is forming around expert judgment both in high tech and in the enterprise.
It demands speed, accuracy, and cognitive load, all having value in the market.
People with early experience who grow with it become super-valuable.
On productivity, history suggests that when a task gets cheaper, it gets used more.
Companies apply AI to more problems, more languages, and more countries, and each launch brings new evaluation needs.
The roles are changing faster than they are disappearing. Our job is to make sure people change with them.
Invezz: AI is expected to bring a lot of benefits to robotics. In your view, how quickly can robotics become sophisticated enough to start doing work that is currently dangerous for humans to do
Robots already do some of this work in repetitive, controlled settings, such as pipeline inspection, assembly plants, warehouse logistics, and handling chemical materials.
Now the next frontier is operating in unpredictable environments. A factory floor follows rules. A farm field, a mine, or a construction site changes every hour.
We worked with an agricultural robotics company to take its weeding system from early experiments to an advanced AI product.
Close to a million images were annotated over the engagement, and the system grew to recognize hundreds of plant types, with clear gains in detection accuracy and in how well it handled new conditions.
Better tooling and automation cut completion time for the most complex scenes by 61 percent, and real-time monitoring let the team spot and correct edge cases quickly.
That last point is the strategic one.
The first 99 percent of training teaches a system what it will usually see.
The last 1 percent is where trust is earned: the weed that looks like a crop seedling, the plant half covered in mud, the field under harsh light.
A robot also has to combine camera, LiDAR, and other sensor data into one consistent picture of the same moment.
Data can look correct on a screen and still fail in the field. Human judgment tells the two apart.
Progress will come workflow by workflow. Narrow, repetitive work in tough conditions will move first, likely within the next few years.
Work that requires a robot to improvise will take longer. Until then, the most reliable setup is a robot doing the physical work with a human ready to step in remotely.
Invezz: India supplies a large share of the world’s annotation workforce. As routine labeling gets automated, what happens to entry-level annotators in smaller cities? What does the career path look like from a basic labeling role to an expert evaluator, and how is iMerit helping people make that move?
Smaller cities are where much of India’s new work is happening.
The Quess Pulse Report for the first half of fiscal 2026 found that 69 percent of payroll-linked employment is in Tier 2 and Tier 3 cities.
Tier 3 towns alone account for 40 percent of the workforce and Tier 2 hubs for 29 percent, compared with 31 percent in Tier 1 cities. The future of AI work in India will be decided there.
Routine labeling is being automated, and entry-level roles are changing with it. Inside iMerit, four kinds of roles are growing:
- Ground truth builders turn images, video, documents, sensor data, and medical scans into data a model can learn from. They also write the prompt and response pairs used to tune a model.
- Model evaluators assess what a model produces, correct it, and track patterns of failure with every update.
- Red teamers test a model in creative ways and document exactly where it fails, often across languages and cultures as AI reaches new countries.
- Human in the loop experts step in once a model is live. When an AI system in insurance, finance, or medicine is uncertain, the decision comes to them.
A new team member now starts by checking and correcting what a model produces.
From there, people specialize in a field such as LiDAR, medical imaging, or a regional language, then move into evaluation, red teaming, and project leadership.
About 3,200 people work with us full-time, and 90 percent of them are in India.
Around that core is a network of more than 30,000 curated experts, our Scholars, who join projects when their expertise is needed.
The full-time team gives us consistency and quality over the long term. The Scholar network gives us reach into almost any field a model needs to learn.
The most important change in India is the kind of work being done.
For a global pharmaceutical company, pathologists in India handled first-pass annotation of tissue slides, and US subspecialists reviewed their work.
Indian talent is moving from processing data to shaping clinical judgment for global medicine. Across every role, expert judgment is what keeps AI safe and reliable.
Invezz: Your Scholars network includes physicians, linguists, scientists, and engineers. As models get better, does the value of human data go up or down, and why?
It goes up. Better models make more convincing mistakes.
A fluent answer to a medical or financial question can still be wrong, and catching it takes a physician or a mathematician with years of practice.
AI development is also shifting.
The focus is moving from building ever more capable foundation models to adapting them for specific uses, especially in regulated industries like healthcare, banking, and insurance.
That work depends on experts who can challenge a model before it reaches customers.
We call it tormenting models. We question the model, question it again to make sure it is working, and actually try to break it.
That is why we built the iMerit Scholars platform on Ango Hub.
Our community includes more than 30,000 vetted experts across 118 countries and 106 languages, and the platform is designed to grow to hundreds of thousands.
Every Scholar goes through a three-tier assessment. The first screens general AI capability and cognitive skills.
The second tests domain qualifications and the ability to work effectively with AI models. The most specialized roles add peer review by other experts in the field.
Being part of EXL gives us access to deeper healthcare and insurance expertise, which strengthens our ability to support AI built for specific enterprises.
Our biggest need right now is to go into new expertise, scale into newer languages, and into newer levels of the same expertise.
The next phase of generative AI will be won by those who treat expertise as human capital. Technology should make that expertise easier to reach, retain, and apply.
Invezz: In areas like healthcare AI and autonomous vehicles, a flawed training set can cause real harm. What does iMerit do to catch bias, errors, and blind spots in data before they reach a model?
Most harmful errors in training data are systematic.
A category is defined poorly, a condition is missing, or reviewers read the same thing in different ways. We focus on catching these patterns early, and the work starts with who checks the data and how.
In healthcare, the standard is clinical truth. Our open-source 3D mammography dataset, released in 2026, covers 558 patients, 271 with malignant findings and 287 with benign ones.
Every case is confirmed by biopsy and reviewed by US board-certified radiologists.
The average tumor is 1.34 cm, and 85 percent of lesions are under 2 cm.
This helps a model learn from early cancers, which are the hardest to detect. On every scan, a radiologist has marked the exact boundary between a lesion and healthy tissue.
That judgment is now part of the data. Fast Company recognized the dataset in its Next Big Things in Tech 2026 list.
In autonomous mobility, the standard is consistency across people, sensors and regions.
On one program covering images, video and 3D LiDAR across the US, Europe and Asia Pacific, LiDAR accuracy was stuck near 80 percent.
Our specialists held weekly calibration sessions so reviewers made the same calls, and added a two-stage quality check.
They also revised the taxonomy, adding classes for road objects the model kept missing and removing ones that added noise. Accuracy rose above the client’s 95 percent threshold.
That’s the main reason clients choose to work with iMerit. In high-stakes situations, you want professionalism, accountability, and consistency.
Invezz: Will AI mostly benefit large labs and wealthy countries, or can it spread opportunity more widely? How can countries like India participate in AI as creators and shapers of models, and not just as a source of low-cost labor?
The foundations of AI, such as compute and capital, are concentrated today. The value of AI, however, depends on how well it understands the people it serves.
That understanding comes from everywhere.
A model becomes truly useful when it understands a patient’s history, a farmer’s soil, or a customer’s language.
That knowledge lives with doctors, agronomists, linguists and engineers around the world.
India also has a long tradition of knowing how to do global business – listening to the customer; understanding the problem; solving the problem.
This is very important in applying high-stakes AI in the enterprise.
India brings a deep pool of scientific and professional talent, 22 scheduled languages and a large population whose needs AI must learn to serve.
The greatest opportunity lies in defining how models are evaluated, building datasets that reflect Indian realities, and developing domain models for healthcare, agriculture, and finance.
Those who set the standards will shape what AI becomes.
I have always believed that true breakthroughs happen at the intersection of people and machines. Countries that invest in both will help shape the future of AI.
Invezz: In 2026, iMerit was acquired by EXL. How does pairing your expert data network and Ango Hub platform with EXL’s broader data and AI services change what you can offer clients? Where do you see the company’s future moving forward?
I see three directions ahead.
First, it brings frontier lab discipline to regulated industries.
EXL’s clients in insurance, banking, and healthcare will need to show regulators how their models were tested, who reviewed them, and how failures were handled.
iMerit has built that rigor through evaluation, red teaming, and RLHF with foundation model companies. Together, we can help enterprises become audit-ready before the rules arrive.
Second, it connects expert judgment to real workflows.
Ango Hub brings data, experts, and quality checks into one place.
EXL knows how claims, underwriting, and operations run every day. Together, we can tune models to a client’s own processes and data, and keep evaluating them once they are live.
That shortens the path from pilot to production.
Third, it widens where we can grow. iMerit’s work in robotics, autonomous mobility, healthcare AI, and physical AI gives EXL a direct role with foundation model and autonomy companies.
EXL’s scale of nearly 68,000 people gives us the depth to take on larger, longer programs.
Both companies were built the same way, with people at the center and domain expertise as the foundation of good AI.
That common ground is why the integration has felt natural from the start.
The post iMerit’s Radha Basu on how AI is reshaping India’s data annotation jobs appeared first on Invezz