The actions of physical AI, including robots, are only as good as the human expertise and data used to build the AI models that train them. ExlService Holdings Inc. this month completed its acquisition of iMerit Technology, a leader in AI model training, evaluation, and reinforcement learning.
Founded in 1999, EXL offers services and solutions to industries including insurance, healthcare, banking and capital markets, retail, communications and media, and energy and infrastructure. The New York-based company has about 68,000 employees worldwide.
Founded in 2012, iMerit provides data annotation for robotics, autonomous mobility, healthcare AI, and other high-tech industries. The San Jose, Calif.-based company said its subject-matter experts and proprietary Ango Hub platform enable customers “to collaborate on complex multimodal data to generate highly curated and validated training artifacts for high-stakes models.”
For instance, iMerit has partnered with Carbon Robotics to help digest millions of plant images needed to create an agricultural model for robotic weeding. Radha Ramaswami Basu, founder and CEO of iMerit, has joined EXL as executive vice president and head of iMerit. She and Rohit Kapoor, chairman and CEO of EXL, replied to the following questions from The Robot Report about the acquisition.
iMerit and EXL strive for more trustworthy models #
What challenges are AI model builders underestimating today, and why?
**Basu: **Models are becoming highly capable day by day. With this comes increasing potential for hallucination or exceeding the scope. The labs, like customers, are starting to realize that the final mile and the concept of trust need work in order to thrive in a critical enterprise process.
A model can perform well on a benchmark and still struggle with edge cases, unfamiliar conditions, or the specialized context of a healthcare or financial workflow. The bottleneck is access to expert, domain-specific data and trusted deployment in a business workflow.
This is solved by experts who can challenge the model, expose failure modes, and evaluate whether its reasoning and behavior are reliable in a given business environment.
How does EXL’s acquisition of iMerit change how models are trained and evaluated?
Kapoor: The acquisition connects parts of the AI lifecycle that have often been managed separately, establishing an end-to-end AI platform for enterprises.
iMerit brings expert-led model training, multimodal evaluation, the Ango Hub platform, and a global network of domain experts. EXL brings enterprise data, deep industry context, and experience integrating technology into business operations.
Together, those capabilities create a more continuous path from preparing proprietary data to fine-tuning a model, evaluating how it behaves, and operationalizing it within an enterprise workflow as clients transition from pilot to production-scale AI. The goal is to help organizations bridge the gap between innovation and real-world production, turning AI potential into measurable business results that are trustworthy.
Physical AI still has to demonstrate business value #
Why aren’t other companies taking your approach?
Kapoor: Enterprise AI is entering a more demanding phase. Organizations have experimented widely, but most still struggle to achieve consistent business value.
The challenge was never just picking a model; it’s training, evaluating, adapting, and governing AI systems to perform reliably in specific business contexts. That’s where capabilities like model evaluation, reinforcement learning, domain-specific data, red teaming, and expert human feedback become critical.
AI economics will also push enterprises toward more specialized, purpose-built models, trained on proprietary data, optimized for their own workflows, and continuously improved based on performance. That makes high-quality data, evaluation, and reinforcement learning foundational capabilities, not peripheral services.
This investment reflects a broader market shift: Value is moving from building AI models to making AI usable, trusted, cost-effective, and outcome-driven.
Human experts help guarantee safety, notes iMerit head #
Can you give examples of how this applies to the development of autonomous vehicles and robotics?
Basu: As AI moves into robotics and autonomous systems at the edge, success depends less on model scale and more on data quality. Physical AI must interpret noisy multimodal inputs, reason in real time, and act safely in unpredictable environments to identify the correct next action, not just the “right” answer.
The challenges evolve from perception to planning and decision-making. Human experts play a critical role in creating high-quality ground truth data, and they help build realistic scenarios, identify rare edge cases, and evaluate whether a vehicle or robot responded appropriately.
In autonomous driving, AI is trained on multimodal data — vision, lidar, audio — and tested through simulated scenarios like collisions. Experts evaluate not just what the vehicle sees, but also how it behaves and explains its decisions, ensuring it can act safely in real-world edge cases. These are complex, high-context, and high-cognition scenarios.
What are the implications for safety and compliance?
Basu: The blocker for AI deployment now is trust, accuracy, and validation versus model capability. Thus, safety and compliance cannot be treated as a final check before launch. They must be designed into the data, training, evaluation, and monitoring process.
Expert-driven data and domain insight can be transformed into interaction patterns that assess, torment, and oversee AI models in mission-critical applications.
In regulated industries, organizations must be able to demonstrate not only that an AI system performs well on average, but that it was tested against the situations where an error could create irreversible clinical, financial, or operational harm. EXL strengthens this process through trace analysis, enabling organizations to follow the data and model behavior that led to a particular decision and better understand where errors or unexpected outcomes originated.
Competitive advantage comes from tech combos, domain expertise says EXL CEO #
How do you see the AI market evolving, both in terms of the market and the applications?
Kapoor: Competitive advantage in enterprise AI won’t come from any single layer. It will come from bringing the right layers together: proprietary data, domain context, model evaluation, AI engineering, and governance.
Foundation models will keep improving, but in the enterprise, the question isn’t which model is most powerful. It’s which AI system delivers the most reliable outcome in a specific workflow, under real regulatory and operational constraints. That’s a fundamentally different challenge.
Over the next three to five years, proprietary enterprise data will become one of the most valuable competitive assets. Companies will increasingly want specialized models trained on their own data, not just access to external model providers. And they’ll need continuous evaluation and reinforcement learning to keep those models accurate and safe as conditions change.
Domain expertise will matter more, not less. In complex workflows, underwriting, claims, healthcare payment integrity, fraud detection, credit decisioning, the model needs to understand the business, the risk, and the regulatory environment. That’s where EXL has a differentiated position.
The companies that win in enterprise AI will be those that combine data, context, AI, evaluation, and execution into one operating model. That has been EXL’s strategy. The iMerit acquisition strengthens it and positions us to play a more central role in how next-generation AI is built, governed, and deployed for real business outcomes.
Editor’s note: Sudeep George, chief technology officer at iMerit, recently participated in The Robot Report‘s webinar on “Physical AI and Robotics.”