For much of the past two years, the artificial intelligence race has been defined by ever-larger models, bigger computing requirements and multibillion-dollar investments in data centres. But a growing number of enterprises are taking a different approach: using smaller, specialised AI models and reserving powerful large language models (LLMs) for tasks where they are genuinely needed.
The trend is being driven by concerns around costs, governance, predictability and data control as organisations move AI projects from experimentation into production environments.
According to Hyther Nizam, CEO of Zoho Middle East and Africa, enterprises are increasingly looking beyond the industry’s fixation on giant frontier models and exploring technologies they can customise and manage themselves. “Large frontier models with trillions of parameters are increasingly becoming commoditised,” he said. “More enterprises are looking toward open-weight models that they can take, customise, fine-tune, and host themselves.”
The shift reflects a broader evolution in enterprise AI adoption. While companies remain eager to use AI to accelerate software development and improve productivity, many are becoming more selective about where the technology should be applied.
Rather than relying on AI for every process, organisations are increasingly using conventional software code for routine workflows and deploying AI only when tasks involve judgement, contextual understanding or content generation. “A principle emerging across the industry is to push more work to code and use AI only where it adds meaningful value,” Nizam said.
The reasoning is practical. Deterministic software produces predictable outcomes, while AI systems are probabilistic and may generate different responses to the same prompt. For heavily regulated industries, predictability and auditability often matter more than automation alone.
Cost is another factor influencing corporate AI strategies. While AI-assisted development can shorten development cycles and speed up prototype creation, organisations are also facing higher spending on governance, security, compliance and testing frameworks. They must additionally account for cloud infrastructure, model access fees and token consumption costs.
Although productivity gains are real, Nizam said they are more measured than some of the industry’s more ambitious claims, with organisations currently seeing “approximately 20% to 30% productivity improvements across the software development lifecycle.”
At the same time, concerns about data sovereignty are becoming increasingly important as AI systems gain access to sensitive corporate information. Businesses want assurance that proprietary data remains under their control and within their jurisdictions, particularly as AI becomes embedded in critical operations.
These concerns are helping fuel interest in smaller, domain-specific models that can run on local infrastructure. According to Nizam, many business applications do not require powerful GPU-intensive systems and can be handled by specialised models designed for narrow use cases.
“Many business problems do not require large GPU-intensive models,” he said. “Enterprises are increasingly deploying specialised models capable of running efficiently on CPUs rather than high-end GPU infrastructure.”
The attraction is clear. Smaller models can reduce infrastructure costs, improve control and lessen dependence on hyperscale cloud providers, while still delivering useful business outcomes.
As enterprises move beyond AI experimentation and focus on real-world deployment, a more nuanced strategy is emerging. Rather than betting exclusively on increasingly larger models, companies are combining open-source AI, specialised domain-specific tools and LLMs where advanced reasoning is required. As Nizam put it, the industry’s direction is becoming “a combination of three strategies: leveraging open-source models, deploying smaller domain-specific models whenever practical, and reserving large language models for tasks requiring advanced reasoning, judgement, or complex generation.”
That approach may ultimately reshape how AI is deployed across the enterprise, shifting the conversation from model size to practical business value.
Source: Khaleej Times