On August 28, EXL chairman and CEO Rohit Kapoor and iMerit founder Radha Ramaswami Basu sat for a joint interview and, for the first time, walked through what the acquisition is actually for. The deal was signed July 6, closed August 3, and is valued at up to $310 million across upfront and future consideration.
The price is not the story. The buyer is.
EXL was founded in 1999, trades on the Nasdaq, is headquartered in New York, and runs roughly 68,000 employees across six continents. Its business is the back office of insurance, healthcare, banking and capital markets, retail, communications and media, and energy and infrastructure. Claims adjudication. Reconciliation. Collections. Document review. That list is close to a complete inventory of the roles AI has cut hardest over the last eighteen months.
iMerit, founded 2012 in San Jose, does model fine-tuning, evaluation, and reinforcement learning. Its asset is not the software. It is a network called Scholars, made up of physicians, scientists, engineers, linguists, and other subject-matter experts, working through a proprietary platform called Ango Hub.
A company that sells labor by the seat bought a company that sells judgment by the case. The gap between those two things is the gap the white-collar labor market is currently trying to climb.
Basu names the bottleneck, and it isn’t the model
Basu’s clearest line in the interview: the blocker on AI deployment now is trust, accuracy, and validation, not model capability.
She described the expert’s job in operational terms. Challenge the model. Expose its failure modes. Evaluate whether its reasoning and behavior hold up inside one specific business environment. A model can clear a benchmark and still break on edge cases, unfamiliar conditions, or the workflow of a particular hospital. Closing that gap requires someone who knows the domain well enough to catch the machine being wrong.
Physical AI makes the point concrete. iMerit processed millions of plant images for Carbon Robotics to train an agricultural model for laser weeding. In autonomous driving, the training input is multimodal — vision, lidar, audio — and the expert grades not just what the vehicle perceived but how it behaved and how it explained the decision. Basu called these high-context, high-cognition scenarios.
Kapoor supplied the other half: value is moving from building AI models to making AI usable, trusted, cost-effective, and outcome-driven.
Read as earnings-call language, that sentence is filler. Read as a labor statement, it says the money that went into pretraining for three years is migrating downstream into evaluation, red-teaming, and domain feedback, and every one of those is headcount-intensive.
The credential floor is rising, not falling
Line this deal up against the rest of 2026 and the shape is unmistakable.
On April 16, Meta terminated a contract and Sama issued redundancy notices to 1,108 workers in Kenya. Their work was bounding boxes, labels, and footage review. General annotation, no credential required.
On August 6, Genpact posted 7.1% revenue growth with headcount down both sequentially and year over year, and non-FTE revenue crossed half the total for the first time. On August 13, Globant carried 2,673 fewer people and watched margin fall anyway.
Now EXL writes a check, and what it buys is physicians, scientists, engineers, and linguists.
Two ends of the same supply chain are moving in opposite directions. The low-credential tier is being cut. The high-credential tier is worth $310 million. Everything in the middle is being squeezed from both sides at once: form-filling, data entry, procedure-following, document routing.
One more date worth holding: on August 18, EXL closed a new $1 billion senior secured credit facility. Deal closes August 3, facility closes August 18, executives explain the strategy August 28. Three events, one month, one direction. The company is loading up to buy again, and what it wants to buy is not capacity. It is credentials.
The job title changed. The building didn’t.
For the people inside this, the change has a specific shape.
The economics of outsourcing did not move. Low-cost labor serving high-cost clients is the same arbitrage it has been since the 1990s. What moved is which hour gets sold.
The old unit was “process this claim.” The new unit is “red-team this claims model and find the policy types where it answers wrong.” Same building, same time zone, same client list, different entry requirement. The first job rewards procedural fluency. The second requires that you actually understand actuarial practice, or clinical coding, or one jurisdiction’s compliance text, well enough to tell when the machine has it backwards.
That gap is bigger than it looks. Being able to judge where a model went wrong is a harder skill than being able to follow the procedure correctly. The first demands that your understanding exceed the model’s. The second only demands that you not make mistakes.
So the roles that expand from here are narrow and specific. Domain evaluators, licensed or equivalently credentialed. Red-teamers who can construct the scenario that breaks the model. Annotation QA, shifting from piece rate to adjudication. And the trace-analysis line Basu described: following the data and model behavior backward to find the step where a decision went wrong. That last one is the closest analogue to traditional audit and risk control, except the subject under review is a model instead of a person.
Contracting: general annotation, basic data entry, and every back-office role whose core competency is “completes the SOP.” These are not being replaced by AI in a single stroke. They are being compressed from above, as the tier above them raises its credential bar, and from below, as the unit price falls past the point where hiring a person pencils.
Kapoor’s line about value moving from building models to using them well translates, in hiring terms, to this: the white-collar dividing line over the next two to three years is not whether you can operate an AI tool. It is whether you own a domain deeply enough to say where the model is wrong. That threshold is precisely what generic upskilling programs cannot teach.
We wrote earlier about Salesforce booking 32 billion “agentic work units” into an earnings deck without a denominator. The EXL deal is the same picture from the other side. The denominator is being redefined, and the people redefining it are the ones getting bought.
Sources
- EXL acquires physical AI model developer iMerit — The Robot Report, 2026-08-28
- EXL completes acquisition of iMerit — GlobeNewswire, 2026-08-03
- EXL to Acquire iMerit — FinSMEs, 2026-07-06
- EXL closes new $1 billion senior secured credit facility — GlobeNewswire, 2026-08-18
- Carbon Robotics partners with iMerit — The Robot Report, 2026-08-26
