Enterprises are increasingly expected to build and own small language models (SLMs) trained on proprietary data rather than rely solely on rented foundation models as they seek greater control over intellectual property and regulatory risks, according to Rohit Kapoor, chairman and chief executive of ExlService Holdings, Inc., a Nasdaq-listed analytics and digital solutions company.
"We're going to see more evolution of small language models and specialized language models that enterprises will want to own. Anytime you're renting a model from somebody else and the regulation changes, your entire business can be put at risk," Kapoor said.
This regulatory anxiety intensified after 12 June, when the US government abruptly ordered Anthropic to block all foreign nationals from accessing its newly released Claude Fable 5 and Claude Mythos 5 models. The unprecedented intervention has accelerated enterprise discussions around data sovereignty, AI supply chain vulnerabilities, and the risks of relying too heavily on foreign AI vendors.
Hybrid AI strategies
Rather than depending on a single technology vendor, businesses are increasingly adopting hybrid AI strategies, in which some capabilities are developed in-house, others are licensed, and some are accessed through external providers, Kapoor said. "The data assets that enterprises have sitting within their own organization are something they don't want to pass on to model companies and lose the IP. The context is becoming the moat. The context is the most important layer overall," he added.
Kapoor cited US insurers such as Travelers and AIG, along with JPMorgan, as examples of organizations that have already begun developing enterprise-specific AI models.
Further, with rising large language model (LLM) token consumption straining enterprise AI budgets, Kapoor said one of the best ways to optimize costs is by building SLMs. "Our belief is that most enterprises will start creating small language models that they train using their own proprietary data assets, and they will own these small language models," he added.
Unlike general-purpose LLMs, SLMs are designed for specific enterprise use cases, such as customer service or document processing. Because they contain fewer parameters, they require less computing power and process fewer tokens, making them significantly cheaper to deploy and operate.
The company recently announced it would acquire AI data solutions company iMerit in a deal valued at up to $310 million. The acquisition is expected to significantly expand EXL's India footprint. iMerit employs around 3,600 people, most of whom are based in India, with a strong presence in Kolkata and operations across tier-II and tier-III towns.
Speaking about India, he said, “As we grow our business, we'll continue adding AI talent in India. Capabilities such as model evaluation, reinforcement learning, and bringing together domain experts for AI training are increasingly being built here.”
Wanted: forward deployed engineers
On the growing demand for forward deployed engineers (FDEs), Kapoor said while AI adoption depends on strong data and models, successful implementation also requires engineers who can tailor these technologies to a client's operating environment. FDEs work within client organisations rather than from their own company's offices, helping adapt AI platforms and products to the specific operational needs of businesses.
"We've been training our own people. We started increasing our investment in hiring from colleges, building up a summer internship programme, and creating this capacity within the company. We'd also be looking at making more acquisitions to bring in consulting talent and forward deployed engineers from Silicon Valley who can help us take these capabilities to our clients," he said.
Beyond technology adoption, AI is also transforming how enterprises are organised, Kapoor said. "What we are seeing is a collapse of operations and IT because if you want to use AI, you have to bring operations, IT and the business together. The business side and the technology side need to be combined. In the past, the CIO (chief information officer) ran technology and the chief operating officer ran operations. But with AI, both have to work together," he said.
For the quarter ended 31 March 2026, EXL reported an increase in revenue to $570.4 million from $501 million a year earlier. Gross profit for the quarter climbed to $222.1 million from $193.3 million a year earlier.