Responsible, Reliable and Cost-effective AI at Scale

AI and intelligent automations are frequently presented as a panacea for operational inefficiencies in healthcare today. As adoption accelerates, so does the financial burden. Recent analysis indicates that health plans expect commercial healthcare costs to rise by 9% next year, with nearly 70% of surveyed plans citing AI-driven billing and documentation tools as a top inflationary trend.

While these tools assist health organizations in many ways, the rapid and often disjointed integration of AI across disparate systems, workflows and departments is creating new expenditures and governance pressures that must be navigated. When the underlying data foundation of a health network  is fragmented, unstructured and unstandardized, even the most sophisticated AI models fail to deliver on their potential, leading to high costs, unpredictable outcomes, and significant risk.

Data First: What Makes Healthcare Data AI-Ready

At Smile, we believe that intelligent AI begins with a data-first approach.

High-quality data that is normalized, structured and standardized, and can be validated offers a trusted foundation for all analytics, automations and AI tools. Without it, AI cannot scale effectively or as expected, with the accuracy required for clinical care delivery or the administration that backs it up. A unified and validated FHIR® data foundation makes AI models usable, defensible and cost-effective, as it standardizes data mapping and transformation.

Our platform, OmniVera, transforms fragmented healthcare information into structured, semantically aligned data. It ingests legacy data from disparate sources and from a variety of formats into FHIR, ensuring data is complete and comprehensive. By prioritizing this foundation, we accelerate deployment, improve model accuracy, and drastically reduce the cost of healthcare AI at scale. 

Safe and Responsible AI: What Does It Take to Counter AI Overtrust?

The primary danger of relying on ‘black box’ AI models (where only inputs and outputs are visible and not the internal decision-making process) is the erosion of trust. There is a growing risk of inevitable instances where AI is trusted beyond its capabilities, potentially leading to real-world harm.

This is why we built our approach to AI on transparency, accountability, and disciplined verification. It is characterized by:

• Human-in-the-Loop: We prioritize structured and disciplined human verification with documented validation paths so  outcomes are clinically verifiable rather than assumed.
• Deterministic Models: Whenever possible, we deploy deterministic models over probabilistic ones, providing a full audit trail that ensures visibility, reliability and repeatable results.
• Accountability: If a system makes an error, we have the ability to identify, correct, and learn from it.
  
  

The Efficiency Equation: Reduce AI Scaling Costs with FHIR & CQL 

A major hidden cost of current AI trends is tokenization. Analyzing massive datasets by running them through LLMs repeatedly is expensive and unsustainable. It requires every data element to be tokenized, resulting in high costs.

A more cost-effective approach is a query-based methodology:

• Tokenize the Question: We use AI to generate the necessary logic, which is a low-cost, one-time activity.

• Execute via CQL: We run that query against a Clinical Quality Language (CQL) layer, which returns only the required data.

Query-based methodologies can only be used on a unified and structured FHIR-first data foundation like Smile OmniVera. The solution unifies, cleans, normalizes, de-duplicates data which is then ready for analytics and AI modelling. Data is then put through a logic layer powered by CQL. CQL is also used to codify medical policies, clinical best practices and quality measures against which the data is run, which is a one-time activity. Executing a CQL algorithm is nominal compared to the cost of asking an AI model to answer questions repeatedly. 

The query-based methodology can also point to why decisions were made, enabling transparency and auditability.

Governance, Trust, and the Path Forward

Scaling AI and automations effectively and unlocking the promise of better care outcomes that these technologies hold, requires organizations to focus on a data-first approach, rather than the hype of a ‘model-first’ one. 

Upon a FHIR-first unified data architecture, efficiency-based and innovation-focused AI models, automations and intelligent analytics can be run reliably, with organizations maintaining the ownership and sovereignty of their data. 

Smile OmniVera: accurate, auditable, affordable data foundation that reliably scales AI & automations