Growing the business with small language models

In the rapidly evolving world of machine learning and artificial intelligence, small language models (SLMs) are emerging as pivotal assets for businesses, offering efficiency, customisability, and on-device processing that reduce costs and latency while enabling real-time, privacy-preserving intelligence.

According to McKinsey's The State of AI in 2025, nearly all organisations are now using AI, with 62% experimenting with AI agents, yet most remain in early stages of scaling, highlighting the need for targeted tools like SLMs to capture enterprise-level value.

For CIOs, developers and researchers in Southeast and East Asia, where regulatory scrutiny and data sovereignty are paramount, SLMs unlock new revenue streams through personalised customer engagement and domain-specific applications.

Gartner predicts that by 2028, 30% of generative AI workloads will shift to domain-specific SLMs on-premises or on-device, up from less than 1% in 2024, transforming AI from a cloud-dependent tool to an edge-powered asset that drives competitive advantage.

McKinsey's survey reinforces this, showing high performers achieving cost and revenue benefits by redesigning workflows around AI, with 64% reporting enhanced innovation.

Enhancing compliance through local AI control

SMLs are revolutionising how organisations manage compliance and data sovereignty, particularly in regulated sectors such as banking. By enabling on-premise deployment, SLMs allow firms to keep sensitive data localised, minimising risks associated with cloud reliance. Gartner highlights that SLMs resolve the tension between powerful AI and data privacy, becoming critical enablers for secure enterprise applications.

Gartner senior director analyst Mike Fang emphasises that sovereignty depends more on open-source or open-weighted models downloadable to local data centres, focusing on self-hosted infrastructure rather than shared clouds.

David Hardoon, global head of AI enablement at Standard Chartered, echoes this, stating: "The beauty of small language models is that they allow for on-premise deployment, enabling organisations to ring-fence sensitive data while still leveraging the immense benefits of generative AI."

He adds that fine-tuning with proprietary datasets ensures compliance with regulations such as GDPR, turning compliance into a competitive edge.

“Imagine AI that truly understands and contextualises all your information — customised, transparent, and secure. This not only cuts costs but fosters trust, and ultimately transforms compliance from a hurdle into a competitive advantage.” David Hardoon

For CIOs in Asia navigating diverse regulatory landscapes—from Singapore's strict data protection laws to China's sovereignty mandates—SLMs provide customised, transparent AI that securely contextualises information, cutting costs and building trust.

McKinsey's findings bolster this, noting that while only 39% of organisations report enterprise-level EBIT impact from AI, high performers prioritise compliance in their initiatives, achieving broader value.

Balancing AI innovation with ironclad security

As AI adoption accelerates, striking a balance between its benefits and robust security is essential. Gartner recommends embedding compliance into product development, forming cross-functional teams for use-case assessments, and prioritising regulatory resilience over efficiency.

This includes establishing regional trust agencies and integrating local talent to adapt to geopolitical mandates.

Fang notes that chief information security officers (CISOs) play a pivotal role in model selection and in testing against prompt injection attacks, especially for open-source SLMs.

Hardoon advocates layered strategies: "embedding ethical guidelines from the start, conducting regular audits, and fostering cross-functional teams for oversight."

He stresses human-AI collaboration to identify vulnerabilities early, likening it to "building a fortress where innovation thrives safely."

At Standard Chartered, investments in talent upskilling and adaptive protocols ensure efficiency doesn't compromise resilience. For developers and researchers in East Asia, where cyber threats are rising, these approaches enable scalable AI deployment. Gartner case studies show that one-third of generative AI implementations achieve double-digit business improvements through smaller, controlled models.

McKinsey expands the narrative, revealing that 80% of high performers set efficiency as an AI objective, while also pursuing growth and innovation, with half redesigning workflows to mitigate risks and maximise security.

The future backbone: Sovereign AI in finance

Sovereign AI models are poised to redefine banking over the next five years, emphasising data control amid geopolitical tensions. Gartner predicts that by 2027, fragmented regulations will cover 50% of global economies, driving US$5 billion in compliance investments, with over 75% of European and Middle Eastern enterprises geopatriating workloads by 2030 to mitigate risks.

"Open-source models, including SLMs, growing rapidly, offering performance close to frontier models while providing cost efficiency and controllability, though commercial models will persist for complex tasks." Mike Fang

In banking, Hardoon envisions sovereign AI as "the backbone of finance, driven by a growing need for data control and resilience." He highlights localised models for regulatory nuances, fraud detection, and personalised services, but stresses responsible governance through specialised talent and partnerships.

For Southeast Asian CIOs, this evolution means leveraging sovereign stacks—nations may spend 1% of GDP on them by 2029—to address cross-border risks. As global banks like Standard Chartered collaborate regionally, sovereign AI democratises access, fostering inclusive growth in markets like Indonesia and Vietnam.

McKinsey's survey lends weight to the idea that while 32% expect workforce reductions due to AI, high performers view it as a tool for innovation, with 64% citing AI-enabled creativity aligning with sovereign models' localised potential.

Navigating challenges: Open-source vs proprietary SLMs

Deploying SLMs involves weighing open-source flexibility against proprietary reliability, and it comes with distinct challenges. Gartner advises internal talent on customising open-source models using techniques such as Low-Rank Adaptation (LoRA), offering cost savings and domain expertise through platforms like Hugging Face.

However, proprietary models provide stability but risk vendor lock-in. Fang warns of transparency issues in undisclosed open-source training datasets, plus context engineering, security guardrails, and alignment between business and technology stacks.

Standard Chartered's Hardoon points out: "deploying open-source SLMs offers flexibility and community-driven innovation, but ensuring security against vulnerabilities and integrating with proprietary systems without intellectual property leaks remains a key challenge."

Proprietary options guarantee performance but come at higher costs, and both require internal expertise for governance. For developers and researchers in Asia, where talent shortages persist, vigilance in updates to open-source projects and transparency in negotiations over proprietary software are crucial. Gartner urges urgency: as AI enters the "Trough of Disillusionment" by 2026, SLM strategies will differentiate leaders.

McKinsey underscores this, noting that nearly two-thirds of organisations are still piloting AI, but high performers—often using agents powered by efficient models like SLMs—report use-case benefits, emphasising the need to overcome scaling hurdles through workflow redesign.

Unlocking local innovation via regional alliances

Regional partnerships are key to tailoring AI solutions to local needs and amplifying impact across diverse Asian contexts. Gartner recommends shifting AI gravity to regional hubs across data, technology, governance, and innovation, with 60% of organisations increasing reliance on local solutions.

Fang describes a move from globalised AI to sovereign boundaries, urging regional trust agencies to handle regulations and cultural values, as in Singapore and Thailand.

Hardoon agrees, stressing that regional partnerships help deepen and further contextualise the use of AI for local needs through co-developing or establishing common standards. He cites Southeast Asian collaborations on agriculture or finance models to address language diversity and accelerate innovation.

For CIOs in East Asia, such alliances— like Volkswagen's with XPENG in China—reduce duplication, share resources, and mitigate risks, turning isolated efforts into equitable, hyper-localised solutions.

McKinsey's insights enhance this perspective, showing that 43% expect no workforce change from AI, suggesting partnerships can support reskilling and innovation in regions like Southeast Asia, where localised AI agents could drive transformation.

Something to think about

SLMs, integrated with sovereign strategies, offer CIOs and application developers in Southeast and East Asia a path to resilient, compliant AI. But the journey towards achieving the promised business value of AI will not be easy.

At the 6th Annual C-Engage Conference in Singapore, 60% of the audience attending the CIO Conference believed that success would come when there is a clear business objective. Backed by McKinsey's survey of widespread AI adoption yet limited scaling, and Gartner's projections on SLM growth, organisations can navigate regulatory complexities, foster innovation, and unlock local value by heeding these expert insights.