Think small to release the targeted power of AI in enterprise operations

The rise of large language models and the pressure to act with them has led many leaders to a frenzy of nut-cracking with some very powerful and expensive sledgehammers. As artificial intelligence (AI) continues to mature, it is essential to recognize the enterprise benefits of a more focused and specialized class of models, which are becoming known as small action models (SAM).

SAMs are emerging as a practical and efficient solution for specific, narrowly defined tasks. Unlike large language models (LLM) or large action models (LAM), which are built for a wide range of applications, SAMs are designed to address particular functions within a defined context, suiting targeted operations.

Doers vs. talkers: How small action models differ from small language models

SAMs and small language models (SLM) are specialized models, each serving different purposes within the AI ecosystem. SLMs focus on understanding and generating text, and SAMs focus on performing specific actions in response to the text.

SAMs are tailored to execute particular actions or manage discrete functions within a broader system. For example, a SAM might automate the reconciliation of financial transactions, focusing specifically on identifying discrepancies in certain types of records.

SLMs, on the other hand, focus exclusively on language-related tasks. They are smaller-scale versions of language models designed to understand, generate, or classify text within a limited scope. A customer service chatbot could be classified as an SLM when its functions are limited to text-based interactions. However, when that same chatbot is designed to perform actions based on those interactions—such as triggering workflows, updating databases, or routing inquiries—it shifts into the realm of SAMs.

Built and priced for purpose: Reduced data and computation demands cut the cost of SAMs

SAMs’ focus on specialized tasks often requires less computational power and data than larger models. To the relief of those wielding the aforementioned expensive sledgehammers, this efficiency translates into cost savings and quicker deployment timelines. For example, in customer service, SAMs can automate the classification of inquiries and direct them to the appropriate department, optimizing workflows and enabling human agents to focus on more complex interactions.

A practical example is Zendesk’s Answer Bot, which automates responses to frequently asked customer questions by leveraging a company’s knowledge base. This targeted approach improves response times and reduces the need for intervention by human customer service representatives. Similarly, XOR AI Recruiter automates initial candidate screening and interview scheduling, streamlining the recruitment process in HR departments.

Limited scope makes SAMs easier to predict, control, and govern

One advantage of SAMs is their reduced risk profile. Because they are designed to operate within a clearly defined scope, SAMs are less likely to encounter the unintended consequences that can arise from using more generalized AI models. This narrow focus makes SAMs easier to govern, as their applications are more predictable and manageable.

In financial services, Feedzai’s real-time fraud detection model offers an example of how we can use SAMs to monitor specific types of transactions for signs of fraud. By focusing on particular transaction patterns, this SAM provides a precise solution that helps mitigate risk without the complexity associated with larger, more generalized models.

Understand the role of SAMs within a broader AI ecosystem

SAMs are not the answer to everything (see Exhibit 1). You must use them as a complementary part of the broader AI ecosystem. While LLMs and LAMs handle a range of complex, multi-domain tasks, SAMs manage specific, high-impact operations that require precision. SAMs can deliver targeted improvements in areas where efficiency and accuracy are crucial.

For instance, a SAM could automate the reconciliation of financial transactions, focusing specifically on identifying discrepancies in certain types of records. This targeted approach complements the broader capabilities of LLMs and LAMs, which might oversee more extensive financial processes.

Exhibit 1: SAMs are part of an AI ecosystem, best used in narrow, well-defined contexts

Source: HFS Research, 2024

SAMs execute tasks but aren’t autonomous—and they don’t learn

SAMS will likely be stitched together in end-to-end processes as multi-agent AI systems emerge. Multi-agent AI systems refer to a network of multiple AI agents working together to achieve complex goals.

Each agent in such systems is typically specialized in a particular function or task, and they interact and coordinate with each other to perform more complex operations than a single AI agent could manage alone. However, while a SAM can autonomously execute tasks within its defined scope, SAMs typically operate with a fixed set of rules or parameters. They do not possess the broader autonomy or learning capabilities often associated with more complex agents in multi-agent systems.

This limitation—their lack of the capability to self-learn—means it is essential that you only deploy SAMs after careful consideration of the effectiveness of your current processes. A period of process reengineering should ensure you do not simply ossify out-of-date practices.

Consider SAMs for targeted ops in customer service, finance, and supply chain

Identify the specific areas within your operations where precision and efficiency are paramount. Customer service, finance, and supply chain operations are common areas where firms should consider deploying SAMs. Evaluate offerings from companies such as OpenAI, Anthropic, and Snorkel AI. Work with expert partners to understand their capabilities and limitations.

Establish governance frameworks tailored to the specific applications of SAMs. SAMs are likely to require a lighter touch than LLMs and LAMs. To realize value faster, apply the relevant level of governance vs. the different levels of risk carried by various types of AI.

Involve multiple departments—such as IT, operations, finance, and customer service—in the planning and deploying of SAMs to ensure effective integration in your end-to-end business processes.

The Bottom Line: Make small action models part of your AI toolkit

Small action models offer a practical and efficient way to address your specific operational challenges. By focusing on targeted tasks, SAMs can deliver measurable improvements in efficiency while controlling both risk and cost. Begin by identifying critical areas for deployment, exploring the available models, and developing a risk-appropriate governance framework. With careful planning, SAMs should be integral to your enterprise’s AI toolkit.

See also:

• How to choose the right LLM for your business challenge
• New tech is disrupting how enterprises consume services
• The HFS definition of enterprise AI
• Use our framework to win as a Generative Enterprise
• Five ways GenAI disrupts the enterprise