The general consensus remains that open weight models are suited for largely high-volume, repeatable workflows where accuracy requirements are clear
With rapid advances in open weight models and growing concerns over AI costs, a new paradigm is emerging in enterprise AI deployments - model routing.IT companies are increasingly experimenting with an extra orchestration layer that mix cutting edge frontier models with cheaper open weight ones to carry out their work.Speaking to businessline industry executives and analysts suggest that IT service firms are moving away from the ‘one-model-fits-all’ to ‘right model for the right task’. Though client deployments remain mostly confidential, they signal a growing adoption of simultaneous multi-model architectures across banking, insurance, telecom, and technology sectors.Srikrishna Ramakarthikeyan (Keech), CEO at Hexaware, told businessline recently that the company relies on a ‘harness’ that does ‘model selection’ on a real time basis. “The capabilities of models are converging, but the cost is diverging. So capability variance is maybe 5 per cent, but the cost difference is 20x. The implication of this is that we have to choose the right part of the task to send to the right model. This is the capability of the harness,” he said.Ramakarthikeyan believes that the same frontier lab can have a suite of multiple models with different cost and capability implications. This makes deciding which model to use for which task an important part of AI optimisation.Accuracy NeedsSridhar Ramaswamy, CEO, Snowflake, said in a blog post that while the industry earlier spent the last several years optimizing the capabilities of individual models, the economics of AI now demand ‘intelligence efficiency’ based on cost, speed and performance.Despite their improving capabilities, the general consensus remains that open weight models are suited for largely high-volume, repeatable workflows where accuracy requirements are clear.Raja Lahiri, Partner and Technology Industry Leader, Grant Thornton Bharat said enterprises have recognised that perhaps 70-80 per cent of AI transactions involve routine tasks where open-weight models can deliver satisfactory outcomes at a fraction of the cost.Major Roadblocks“The tasks include things like document summarisation, invoices and claims, customer-service triage, classification and tagging, template-based report generation, translation, and basic coding assistance among others. Frontier models, meanwhile, remain better suited to advanced reasoning, multi-step problem-solving and high-value decision support,” he said.Lahiri believes that for IT services firms, model routing helps in shifting from simply providing model access to designing the architecture surrounding the models.Despite the cost benefits, a key roadblock in model routing is restrictions regarding movement of proprietary data into open weight models without client sign off.Gaurav Vasu, Founder and CEO, UnearthInsight says that confidential and contextual enterprise workloads, such as financial analysis, financial analytics models and people-related models, are not shifting because the underlying data cannot simply move to open-weight models.“IT services economics and delivery are governed by client guidelines and contractual clauses, which restrict experimentation with different models without the necessary approvals,” he said.Published on September 3, 2026
