EXL Buys AI Training Firm iMerit to Close the Gap Between AI Demos and Real Business Results

Key points

• EXL Service Holdings, a New York company with roughly 68,000 employees, completed its acquisition of iMerit Technology in July 2025.
• IMerit, founded in 2012 and based in San Jose, California, specialises in labelling and checking the data used to train AI models for robotics, healthcare, and autonomous vehicles.
• The combined company plans to offer enterprises a single path from raw data to a finished, tested AI system ready for production use.
• EXL CEO Rohit Kapoor told The Robot Report that competitive advantage in AI will come from proprietary data and continuous evaluation, not from picking the most powerful general model.
• IMerit founder Radha Ramaswami Basu joins EXL as executive vice president and head of iMerit.

Most AI projects fail not because the underlying model is weak, but because nobody prepared it properly for the messy, specific conditions of a real job. That's the problem EXL Service Holdings is spending real money to solve.

EXL, which sells technology-driven services to insurers, banks, and healthcare and retail firms, completed its acquisition of iMerit Technology this month. The price wasn't disclosed.

What does iMerit actually do?

IMerit trains and checks AI models. Its subject-matter experts label raw data, the images, sensor readings, and audio that an AI model learns from, then stress-test the model to find where it gets things wrong.

Before a self-driving car can recognise a child running into the road, thousands of human reviewers must watch video footage and mark every child, every kerb, every unexpected obstacle in every frame. IMerit does that work at scale. It also built Ango Hub, a software platform that lets clients and domain experts collaborate on sorting and validating complex data sets. We reported on 27 August how Carbon Robotics recently rebuilt its entire approach to plant recognition, and iMerit's partnership with that company, processing millions of plant photographs so a robot could learn to weed farmland without pulling up crops, is a clean illustration of what that preparation actually involves.

Why does this deal matter for businesses using AI?

It matters because most companies are stuck between a promising AI pilot and a production system they can actually trust. That gap costs money and time, and it's where most enterprise AI budgets quietly disappear.

Kapoor told The Robot Report that value is moving away from building the raw AI model toward making AI usable and consistently accurate inside a specific business workflow. The combined EXL-iMerit operation aims to cover the full chain: gathering proprietary company data, fine-tuning a model on that data, evaluating how it behaves under pressure, then running it inside a live business process.

Basu put the core problem plainly to The Robot Report. A model can score well on a standard benchmark, the kind of exam used to rank AI systems, and still fall apart when it meets the unusual or high-stakes situations that healthcare or financial workflows throw at it daily. Human experts who can challenge the model and identify its failure points are the fix, not an optional extra.

What about robots and self-driving vehicles?

Physical AI, meaning AI that controls a body in the real world, whether a surgical robot or a lorry, has even less room for error than a chatbot giving bad advice.

Basu explained that autonomous vehicles are trained on layered sensor data: camera vision, lidar (a laser-based distance-measuring system), and audio. Human reviewers then evaluate not just what the vehicle detected, but whether its decision and its explanation of that decision would have been safe in a real collision scenario. That review process is exactly where iMerit's experts operate.

Safety and compliance, Basu said, can't be a final checkbox before launch. They have to be built into data collection, training, evaluation, and ongoing monitoring from the start.

What is the honest takeaway here?

The expensive part of enterprise AI is rarely the model itself. That's the thing worth saying clearly here. Preparation and quality control are where projects succeed or fail, and where the spending actually ends up.

If your company is considering deploying AI in a regulated or high-stakes setting, the most useful question to ask any vendor isn't "which model do you use?" but "who checks whether it's actually right, and how often?"