SLMs vs. LLMs: Efficiency and adaptability take centre stage

In an exclusive interview, Balakrishna D. R. (Bali), Executive Vice President, Global Services Head, AI and Industry Verticals at Infosys, sheds light on the transformative impact of Small Language Models (SLMs) on the artificial intelligence landscape. Bali emphasises how SLMs are delivering specialised, cost-effective solutions tailored to enterprise needs, often outperforming larger models in specific scenarios due to their lower computational and data requirements, higher accuracy, and better domain adaptation.

What are the key differences between Small Language Models (SLMs) and Large Language Models (LLMs), and why are SLMs becoming central to the development of Agentic AI systems?

Small Language Models (SLMs) are transforming the AI landscape by providing specialised, cost-effective solutions tailored to enterprise needs. With lower computational and data requirements, SLMs often deliver higher accuracy and better domain adaptation than large language models (LLMs) in many scenarios. This enables diverse applications, such as bringing intelligence to resource-constrained environments like the edge and addressing concerns around data security and privacy in regulated industries.

SLMs are becoming central to Agentic AI systems due to their inherent efficiency and adaptability. Agentic AI systems typically involve multiple autonomous agents that collaborate on complex, multi-step tasks and interact with environments. Fine-tuning methods like Reinforcement Learning (RL) effectively imbue SLMs with task-specific knowledge and external tool-use capabilities, which are crucial for agentic operations. This enables SLMs to be efficiently deployed for real-time interactions and adaptive workflow automation, overcoming the prohibitive costs and latency often associated with larger models in agentic contexts. Ultimately, they offer a practical path for developing scalable and responsive AI agents.

Could you elaborate on how SLMs enable real-time reasoning and decision-making in dynamic, resource-constrained enterprise environments?

Small Language Models are efficiently designed to run quickly on local edge devices by compressing advanced AI capabilities into compact packages that fit within limited computing resources. By focusing only on the most relevant data, such as real-time sensor inputs or system logs, they reduce computational load without sacrificing context or accuracy. Techniques like early exit decoding allow the model to conclude processing as soon as a confident result is reached, enabling the hardware to handle multiple requests efficiently. Additionally, low-bit quantisation minimises power consumption while maintaining performance comparable to much larger models. Operating entirely on-premises ensures that decisions are made instantly at the data source, eliminating network delays and safeguarding sensitive information. This enables timely interpretation of equipment alerts, detection of inventory issues, and real-time workflow adjustments, supporting faster and more secure enterprise operations.

SLMs also enable real-time reasoning and decision-making through advanced fine-tuning, especially Reinforcement Learning. RL allows SLMs to learn from verifiable rewards, teaching them to reason through complex problems, choose optimal paths, and effectively use external tools. This process "bakes" feedback directly into the model, overcoming the "brittleness" of prompt-only approaches and enabling adaptive, multi-step operations in dynamic environments.

In what ways do SLMs help Infosys and its clients achieve both cost efficiency and sustainability, especially when deployed at the edge or in high-compliance scenarios?

Small language models help Infosys and its clients achieve significant cost efficiency and sustainability by delivering high-performance AI with substantially lower computational and energy requirements compared to large models. Their compact size enables deployment at the edge and in hybrid cloud environments, reducing latency and infrastructure costs while minimising power consumption and carbon footprint. In high-compliance industry sectors, SLMs facilitate on-premise or private cloud deployment, ensuring data privacy and regulatory adherence without compromising performance. This combination of efficiency, flexibility, and compliance makes SLMs a sustainable and cost-effective choice for enterprises aiming to scale AI responsibly and economically.

How is Infosys integrating SLMs into the Topaz platform to support intelligent automation across verticals? Can you share a few specific use cases?

At Infosys, we are leveraging AI across extensive areas. We have launched the natively built small language models, and we are using them in IT operations, cybersecurity, and banking by integrating them into our existing platforms and operations. We have also harnessed AI-driven tools like autonomous robots in warehousing and innovative accessibility platforms to create inclusive experiences. AI has been instrumental in enhancing productivity by automating repetitive tasks, streamlining processes, and accelerating modernisation efforts. This approach is resonating well in the industry, especially for those looking to reduce long-term AI operational costs, explore new deployment models, and enhance risk mitigation.

At Infosys, we are weaving artificial intelligence into the fabric of our operations, from IT management and cybersecurity to core banking services. Our home-grown small language models sit at the heart of this effort, quietly powering help-desk triage, threat analysis, and customer queries inside the platforms we already run. In our finance and accounting operations, we have built agents which work in the background to perform invoice processing and small language models play a key role in some of the tasks executed by agents. Some of these tasks done by the agents are high token consumption in nature and require faster processing. This is one key area where we found SLMs not only reduced the cost exponentially, but also increased the processing speed of the agents.

In our warehouses, autonomous robots coordinate picking and packing in real time, while AI-enabled accessibility tools help us deliver more inclusive digital experiences. This blend of automation and insight is lifting productivity, trimming the repetitive work that slows teams down, and speeding up modernisation projects. Clients appreciate how the approach lowers long-term operating costs, opens flexible deployment options, and tightens risk controls, making AI adoption both practical and sustainable.

With increasing regulatory focus on AI (both in India and globally), how is Infosys aligning its SLM initiatives to ensure compliance, transparency, and ethical AI use?

Infosys is proactively aligning its small language models initiatives with global regulatory priorities by fostering compliance, transparency, and ethical AI practices. Our focus on responsible AI incorporates red-teaming protocols and behavioural testing to identify and mitigate vulnerabilities, enhancing model robustness and safety. Complementing these, we employ automated reasoning techniques like fact verification, consistency checks, and self-refinement loops to reduce AI hallucinations. Governance frameworks, real-time monitoring, and AI audits ensure accountability in high-risk sectors such as BFSI and healthcare. Supported by Responsible AI Guardrails, human-in-the-loop validation, and explainability, Infosys remains committed to deploying reliable, trustworthy AI that prioritises consumer protection.

Infosys integrates regulatory compliance, transparency, and ethical AI directly into the small language models through rigorous safety protocols (red-teaming, behavioural testing), governance frameworks (real-time monitoring, sector-specific audits), and Responsible AI guardrails. We employ automated reasoning for fact verification, human-in-the-loop validation, and explainability to minimise hallucinations, ensure accountability, and prioritise consumer protection – aligning with global standards like India’s DPDP Act and the EU AI Act.

As SLMs and Agentic AI grow in enterprise adoption, what skills or capabilities are becoming essential in the IT services workforce? How is Infosys preparing its talent pool?

To meet evolving industry demands, we emphasise upskilling. For example, we foster internal talent through focused training programs, addressing both specialised areas. Our AI-first initiatives ensure all employees are AI-aware, equipped with foundational AI knowledge. Advanced programs like AI Builder and AI Master further deepen expertise, helping teams to innovate effectively. Looking ahead, we plan to expand AI-driven solutions across industries, creating more advanced capabilities. In future, we will likely have human workers doing higher-order tasks, digital workers (bots/tools) doing deterministic & repetitive tasks, and AI workers, powered by AI agents, doing narrow cognitive tasks. With AI bringing in a higher level of automation, new job roles (in AI) will require humans to excel in problem-solving, adopt first-principle thinking, and quickly adapt to diverse technologies to drive innovation.

We treat upskilling as a strategic discipline. Immersive academies build deep domain skills, while an AI-first curriculum gives every employee a working grasp of machine intelligence. Progression tracks such as AI Builder and AI Master turn that foundation into specialist expertise that feeds daily innovation. As we widen our AI portfolio, the workforce will evolve into three complementary tiers: human talent handling creative and strategic decisions, digital bots executing routine deterministic tasks, and AI agents performing focused analytical work. This shift will raise the bar for human roles, rewarding problem-solving, first-principles thinking, and the agility to fuse new technologies into fresh value.

As AI agents and agentic AI gain traction, IT services need talent skilled in designing and orchestrating complex multi-agent workflows. Essential capabilities include integrating diverse tools, curating high-quality data, and leveraging advanced techniques like Reinforcement Learning to build adaptive, reliable AI systems. The provided sources do not contain information on Infosys's specific talent preparation.

How do you foresee the evolution of SLMs over the next 2–3 years? What developments should enterprises prepare for in the era of fully autonomous AI systems?

Looking ahead, we see a growing trend of leveraging the joint synergy of both SLMs and LLMs to strike a balance between broad knowledge, specialised expertise, and cost effectiveness. Additionally, SLMs are playing a key role in augmenting Agentic AI applications in business process automation, customer operations and software development.

Over the next 2-3 years, small language models will become specialised components within fully autonomous AI agent systems, driven by advanced Reinforcement Learning and distillation. Enterprises must prepare for designing and orchestrating complex multi-agent workflows and robust tool integration. This crucially involves building and curating data within dynamic environments that provide real-time, grounded feedback for RL-based training, moving beyond static datasets for truly adaptive AI.