A pragmatic case for Small Language Models in clinical and operational healthcare environments
The Current AI Narrative in Healthcare
Artificial Intelligence has rapidly become the dominant narrative in healthcare transformation. From clinical decision support to administrative automation, the promise of AI—and particularly Large Language Models (LLMs)—has captured the imagination of executives, boards, and clinicians alike.
The appeal is understandable. LLMs like GPT-4, Claude, and Gemini have demonstrated remarkable capabilities in natural language understanding, content generation, and complex reasoning. Healthcare organisations are rightly excited about the potential for these technologies to enhance patient care, reduce administrative burden, and unlock insights from vast clinical datasets.
The Challenge
However, in the rush to adopt cutting-edge AI, a critical question is often overlooked: Is an LLM actually the right tool for the job? In many healthcare contexts, the answer may well be no—and the consequences of getting this wrong extend far beyond wasted investment.
This article argues for a more balanced, intentional approach to AI adoption in healthcare—one that recognises the strategic value of Small Language Models (SLMs) alongside their larger counterparts, and grounds technology selection in clinical safety, governance readiness, and measurable outcomes rather than hype.
Defining the Difference: LLMs vs SLMs
Before evaluating which model type suits a given healthcare use case, it’s essential to understand what distinguishes Large Language Models from Small Language Models—not just technically, but in terms of their practical implications for healthcare operations.
Large Language Models (LLMs)
LLMs are trained on massive, general-purpose datasets—often spanning significant portions of the publicly available internet. Models like GPT-4, Claude, and Google’s Gemini contain hundreds of billions (or even trillions) of parameters, enabling them to perform a wide range of tasks from creative writing to complex reasoning.
Their strength lies in breadth: they can handle open-ended questions, synthesise information across domains, and generate human-like text. However, this generality comes at a cost—both computationally and in terms of control.
Small Language Models (SLMs)
SLMs, by contrast, are designed for focus. With parameter counts typically in the millions to low billions, models like Microsoft’s Phi-3, Meta’s LLaMA variants, or Mistral 7B are trained on curated, often domain-specific datasets. They sacrifice some of the broad generality of LLMs in favour of efficiency, interpretability, and targeted performance.
For healthcare organisations, this trade-off often works in their favour. Many clinical and operational tasks don’t require the vast general knowledge of an LLM—they require precise, reliable performance within a well-defined scope.
|
Aspect |
Large Language Models (LLMs) |
Small Language Models (SLMs) |
|
Model Size |
Billions to trillions of parameters (e.g., GPT-4, Claude, Gemini) |
Millions to low billions of parameters (e.g., Phi-3, Mistral 7B, LLaMA) |
|
Training Data |
Vast, general-purpose datasets spanning the internet |
Focused, domain-specific or curated datasets |
|
Compute Requirements |
High GPU/TPU demands; typically cloud-hosted |
Lower compute needs; can run on-premise or edge devices |
|
Deployment |
Predominantly cloud-based via third-party APIs |
Flexible: cloud, on-premise, or sovereign infrastructure |
|
Explainability |
Often opaque; difficult to trace reasoning |
More interpretable; easier to validate outputs |
|
Cost |
Higher operational costs at scale |
Lower total cost of ownership for targeted use cases |
The Risks of Over-Indexing on LLMs in Healthcare
The enthusiasm for LLMs in healthcare is not without foundation—but it has, in many cases, outpaced careful consideration of the risks involved. For Australian healthcare organisations navigating complex regulatory environments, tight budgets, and heightened clinical governance requirements, these risks demand serious attention.
Data Privacy and Sovereignty
Most leading LLMs are accessed via cloud-based APIs operated by international vendors. When patient data or clinical information is processed through these services, questions arise about data residency, cross-border transfers, and compliance with Australian privacy legislation including the Privacy Act 1988 and state-based health records legislation.
For public health services and those handling sensitive patient information, the lack of sovereign infrastructure options for many LLM platforms presents a genuine governance challenge.
Clinical Safety and Explainability
LLMs are fundamentally probabilistic systems. They generate outputs based on statistical patterns learned during training—not through deterministic logic or verified clinical reasoning. This creates inherent risks when applied to clinical decision support, diagnostic triage, or any context where patient safety depends on the accuracy and traceability of AI outputs.
Critical Consideration
In a clinical governance context, the inability to fully explain why an LLM produced a particular recommendation—or to reliably reproduce that output—poses significant challenges for validation, audit, and accountability.
Cost and Infrastructure Burden
Running LLMs at scale is expensive. Whether through API usage fees or the infrastructure required to host large models internally, the total cost of ownership can quickly become prohibitive—particularly for healthcare organisations operating under funding pressures and competing capital priorities.
The environmental sustainability of large-scale LLM inference is also increasingly scrutinised, adding reputational and ESG considerations to the financial equation.
Vendor Lock-in and Operational Fragility
Dependence on a single LLM provider creates strategic risk. API changes, pricing adjustments, or service discontinuation can disrupt operations at short notice. For mission-critical healthcare applications, this fragility is difficult to accept.
The Strategic Value of Small Language Models
Against this backdrop, Small Language Models emerge as a compelling alternative—not as a replacement for LLMs, but as a more appropriate choice for many healthcare contexts.
Control & Transparency
SLMs can be hosted on-premise or within sovereign cloud environments, giving organisations direct control over data residency and processing.
Clinical Validation
Smaller models are easier to validate, test, and audit—critical requirements for clinical governance and regulatory compliance.
Lower Total Cost
Reduced compute requirements translate to lower infrastructure costs and more sustainable operational models.
Purpose-Built Performance
SLMs can be fine-tuned for specific clinical or operational tasks, often outperforming larger models within their defined scope.
Perhaps most importantly, SLMs align with the principle of proportionality in healthcare technology adoption: deploying the minimum viable complexity required to achieve a defined outcome, rather than defaulting to the most powerful—and most risky—option available.
Comparative Use Cases in Healthcare
Understanding where each model type delivers the most value requires moving beyond abstract capabilities to concrete healthcare scenarios.
SLM-Led Use Cases
- Clinical pathway decision support with defined logic
- Medical coding, classification, and structured documentation
- Incident, risk, and quality event analysis
- Operational workflow optimisation
- Command centre and operational intelligence dashboards
- Medication reconciliation and allergy checking
- Theatre scheduling and resource allocation
LLM-Led Use Cases
- Patient-facing education and communication
- Research synthesis and literature review
- Clinician copilots for broad knowledge access
- Natural language summarisation across large datasets
- Policy drafting and document generation
- Complex, open-ended clinical queries
- Multi-modal analysis (combining text, images, data)
Key Insight
The distinction is not about capability alone—it’s about fit. Many high-value healthcare use cases are inherently narrow and well-defined. For these, an SLM is not a compromise; it’s often the superior choice.
The Right Question: ‘What Problem Are We Solving?’
The most important shift healthcare leaders can make is to reframe AI adoption away from technology selection and toward problem definition. Before asking “Should we use an LLM?”, organisations should be asking:
- 1. Clinical Safety: What are the patient safety implications if this AI system produces an incorrect output?
- 2. Measurable Outcomes: How will we measure success, and what evidence threshold do we require?
- 3. Governance Readiness: Do we have the policies, validation processes, and oversight structures in place?
- 4. Workforce Adoption: Will clinicians and staff trust and effectively use this tool?
- 5. Long-term Sustainability: Can we afford to operate, maintain, and evolve this solution over time?
When these questions are answered honestly, the technology choice often becomes clearer—and frequently points toward more targeted, governable solutions than the largest available model.
The 9Points Perspective
At 9Points, we take an independent, vendor-agnostic approach to AI strategy in healthcare. We believe that technology selection should be driven by outcomes—not by vendor roadmaps, industry hype, or the fear of missing out.
Our experience working with Australian public and private healthcare organisations has consistently shown that the most successful AI implementations are those grounded in:
- Fit-for-purpose technology selection—matching the tool to the task, not the other way around
- Robust clinical and data governance—ensuring safety, accountability, and regulatory compliance
- Commercial realism—delivering value within the financial constraints of healthcare
- Balanced AI strategies—combining SLMs, LLMs, and traditional analytics where each adds the most value
Conclusion: From AI Hype to Healthcare Impact
The conversation about AI in healthcare has, for too long, been dominated by the assumption that bigger is better. Large Language Models are remarkable technologies—but they are not a universal solution. In many healthcare contexts, Small Language Models offer a better balance of performance, safety, governability, and cost.
As Australian healthcare organisations navigate increasing pressure to digitise and innovate, the imperative is not to adopt AI fastest, but to adopt it wisely. That means:
- Recognising that SLMs deserve greater attention for targeted, high-value use cases
- Understanding that LLMs still matter—but should be deployed deliberately, for the right problems
- Embracing intentional, governed AI adoption that prioritises safety and outcomes over novelty
The question is not “Which model is most powerful?” but “Which model is most appropriate for the outcome we’re trying to achieve?”
Healthcare leaders who ask this question—and answer it honestly—will be best positioned to realise the genuine benefits of AI while protecting patients, staff, and their organisations from unnecessary risk.
About 9Points
9Points is an independent Australian advisory firm supporting healthcare organisations with strategy, technology, data, and commercial growth. We help organisations design AI strategies that deliver real impact—grounded in governance, commercial realism, and clinical outcomes.
Ready to Rethink Your AI Strategy?
9Points helps healthcare organisations design balanced AI strategies that combine SLMs, LLMs, and traditional analytics—grounded in governance, commercial realism, and clinical outcomes.