Healthcare IT Today Podcast: 4 Opportunities to Ease the Tension Between Payers & Providers

When it comes to providers and payers, there’s no avoiding the tension that exists between the two. Ultimately, one’s revenue is the other’s costs. There’s also the fact that providers and payers are optimizing for different things. Providers want to ensure patients get the best care. Individual clinicians are incentivized to provide more care at the individual level to serve the patient and avoid malpractice suits; at the institutional level, more procedures mean more revenue, typically. Payers, on the other hand, face their own competitive dynamics as they sell to employers and individuals who want low premiums above all else. 

You might be surprised to learn that despite the inevitably of this tension, there also lies plenty of opportunity in the space between providers and payers. In a recent episode on Healthcare IT Today Interviews, Steve Rowe, Healthcare Industry Lead at 3Pillar, and host John Lynn discuss why this tension exists and what can be done about it. We’ve captured the four biggest opportunities below.

The first opportunity is around Revenue Cycle Management (RCM) and claims. Payers have all sorts of different rules around what they will approve and what they’ll deny to balance the tension between keeping premiums low and paying for medical coverage. These rules are sometimes even group-specific. 

The challenge? Providers don’t know what those rules are, which creates difficulties for the member. It’s not easy to understand at the moment what will be approved and what will be denied. That means patients may end up unhappy when a proposed treatment is denied or not paid in full (and they are balanced billed).

The opportunity here is for payers to expose that logic to health systems—essentially preadjucating payment (instead of doing it after the fact). The business rationale: make it easy for in-network providers to get paid in exchange for more competitive rates. Some companies are already doing this:  “Glen Tullman is doing it with Transcarent; he’s essentially trying to intermediate the payer to create a new network. His whole premise to providers is, ‘Join our network because we will pay you the same day you do service,’” notes Steve. “That’s how he’s building his network with the top healthsystem and doctors.”  

In Steve’s experience building an RCM startup and working with a regional Urgent Care chain, he observed that the expertise and institutional knowledge around claims processing was largely in the heads of the medical billing coders. 

There are two main forms of complexity in RCM he highlights:

1. Submitting the correct eligibility information (e.g. specific formatting of member ID numbers)
2. Matching the right diagnostic codes to the appropriate CPT codes, which can be a large and complex matrix.

The risk here is that this institutional knowledge will be lost when these experienced medical billers retire. The processes are very manual, with reimbursements not keeping up with labor inflation. 3Pillar is leveraging AI and data mining to reverse engineer each payer’s algorithm for approvals and denials. The goal is to systematize this knowledge and flag issues proactively, rather than relying on the institutional knowledge of the billing staff.

The vision is to integrate this RCM intelligence engine with clinical documentation tools. That way providers are alerted in real-time during the care planning process about treatments or codes that are likely to be denied by the payer. This will improve the financial experience for providers and patients alike.

There is a significant opportunity for data transformation as regional payers have data that lives in separate systems that don’t talk to each other. The pipes to connect these systems haven’t been built and the data isn’t defined in the same way. Regional payers are often at a technological disadvantage compared to national payers because they still have on-premise servers and haven’t moved to the cloud. The IT departments for these payers are swamped putting out fires. They simply don’t have the resources to take on the work associated with major technology modernization projects.

And here’s the rub: Self-insured employers want highly customized insurance products and plans that require flexible and configurable technology platforms. National payers have invested in modern tech stacks that can support this level of customization. However, regional payers struggle to match this same capability. 

So, there’s a real need for regional payers to create a unified data platform and operating system that can integrate data from various systems (e.g., claims, population health, PBM, etc.). This would result in a simplified member experience while enabling seamless workflows for call center representatives, who often have to navigate multiple disparate systems. This is an area where working with a partner who specializes in this capability would be beneficial. 

Speaking of member experience, it’s now the number one concern of Vice Presidents of Benefits at self-insured employers thanks to a tight labor market. Top-tier benefits are necessary to attract and retain talent. There’s no doubt that there’s plenty of room to improve. The experience is often fragmented and frustrating as members struggle to get accurate information about coverage, costs, and provider networks. 

There’s an opportunity for payers to make their medical policies and coverage algorithms more transparent and accessible to members at the point of care. Steve explains, “I’m excited about this opportunity because we’ve all been there where it’s like, ‘I just want to know if this particular provider for urgent care who is still open at 10 p.m. is in network. I can’t figure that out on the app. There’s not a search function and the call line doesn’t open until 8 a.m. tomorrow.”

What if patients could get real-time answers to their questions? 3Pillar is making that vision a reality through chatbots powered by AI and knowledge graphs. By using AI to combine data from disparate systems, members can get accurate, up-to-date information at any time, from anywhere.

These chatbots could also help to address the challenge of call center representatives needing to navigate multiple systems to piece together an answer for a member. Steve points out one key consideration: ensuring the chatbots are fed accurate data and avoiding hallucinations. Doing so requires careful design and integration with the underlying data sources. 

While none of these opportunities have “easy buttons” to press, they all provide means for payers to differentiate themselves and better serve patients and providers. You can discover even more areas for payers and providers to win in the full podcast episode. 

The Importance of Guardrails for Secure and Responsible AI Applications

Artificial Intelligence (AI) and Large Language Models (LLMs) are reshaping industries by enabling smarter, more intuitive interactions with technology. However, with great power comes great responsibility. Without appropriate safeguards, these applications can produce unpredictable outputs, leak sensitive information, or even amplify harmful content. This is where guardrails come into play.

Guardrails are an essential layer of defense, ensuring AI systems remain safe, reliable, and aligned with ethical guidelines. They not only enhance security but also optimize performance, enabling organizations to strike the right balance between innovation and risk mitigation.

In AI applications, guardrails refer to a collection of frameworks, processes, and tools designed to monitor and regulate system behavior. These mechanisms proactively prevent unintended or harmful outcomes by:

• Blocking malicious or inappropriate inputs.
• Filtering harmful or false outputs.
• Detecting vulnerabilities, such as prompt injections or hallucinations.
• Protecting sensitive or proprietary data from leaks.

Guardrails are especially critical for LLM-powered applications, such as chatbots and virtual assistants, where user trust is paramount. They ensure that while AI delivers accurate, efficient, and innovative responses, it also remains secure, ethical, and compliant.

Not all guardrails are created equal, and their effectiveness depends on the application’s specific needs. Below are the primary types of guardrails and their unique advantages:

The simplest and fastest method, rule-based guardrails rely on predefined criteria, such as regular expressions or keyword lists, to block or validate content. For example, a profanity filter might block offensive language, while a format validator ensures inputs meet specific structural requirements. Though straightforward, this approach is limited in handling nuanced or context-dependent issues.

These guardrails leverage LLMs themselves to assess the coherence, relevance, or alignment of inputs and outputs. Metrics like perplexity (measuring word coherence) or alignment scores (ensuring outputs match guidelines) help detect deeper semantic issues. This approach is ideal for applications requiring a sophisticated understanding of language patterns but may involve higher latency.

More advanced than metrics, LLM judges are models specifically trained to assess and validate content. They can identify toxic language, verify factual accuracy, or evaluate responses against specific criteria. While powerful, their reliance on multiple LLM calls can increase latency and costs.

By designing prompts that guide the AI’s behavior, developers can reduce the likelihood of generating harmful or irrelevant content. For example, prompts can instruct the model to avoid answering personal or inappropriate questions. Chain-of-thought (CoT) techniques further enhance precision by structuring prompts with step-by-step instructions and examples.

The growing need for robust guardrails has spurred the development of tools and frameworks. Here are some leading solutions:

• Aporia: A real-time platform for mitigating LLM hallucinations, inappropriate responses, and prompt injections, complete with pre-made policies and dashboards.
• NeMo Guardrails: An open-source toolkit by NVIDIA, offering customizable input, dialog, and output safeguards.
• Guardrails AI: A flexible Python framework for validating inputs and outputs, enabling tailored guardrails for any AI application.
• Azure Guardrails: Microsoft Azure’s built-in safety features, providing prompt injection shields, sensitive data filters, and groundedness detection.

These tools simplify the implementation of guardrails, enabling developers to focus on creating impactful applications without compromising safety.

According to the Open Worldwide Application Security Project (OWASP), AI systems face unique vulnerabilities. Here is how guardrails provide essential protection against some of the risks pointed out by OWASP’s security study:

Prevent attackers from injecting harmful instructions into AI inputs using prompt injection shields. These guardrails block malicious user prompts before they reach the AI model.

Guardrails can validate outputs to ensure they don’t trigger unsafe downstream processes, such as unauthorized database queries or code execution.

Filters are able to detect and redact personal or proprietary information from AI responses, ensuring compliance with privacy regulations.

Hallucination detectors validate the factual accuracy of outputs by cross-referencing trusted data sources, reducing the risk of misinformation.

Guardrails can be made to limit the scope of actions AI can take autonomously, preventing unintended consequences due to excessive permissions or autonomy.

By addressing these challenges, guardrails ensure AI applications remain secure and trustworthy.

The “ideal” set of guardrails depends on several factors, including application type, user expectations, and budget constraints. For instance:

• Real-time chatbots require faster, rule-based solutions to minimize latency.
• Applications processing sensitive data may prioritize advanced LLM-based safeguards.
• Organizations with limited budgets might adopt open-source frameworks like Guardrails AI or NeMo.

Moreover, asynchronous guardrails—where validation occurs parallel to output delivery—can enhance speed without sacrificing security. This flexibility allows organizations to customize their approach, ensuring guardrails align with their unique objectives.

As AI continues to transform industries, guardrails are no longer optional—they’re a necessity. These safeguards not only protect users and organizations from risks but also reinforce trust, paving the way for more responsible AI adoption.

By integrating the right combination of rule-based, AI-powered, and engineered safeguards, developers can build applications that are secure, ethical, and effective. The journey toward responsible AI begins with the right guardrails in place.

If you would like to explore how our Innovation Lab can help you implement effective AI guardrails, don’t hesitate to reach out. Our team is here to guide you in building secure, ethical, and high-performing AI solutions.

Jakub Mlady

Senior Software Engineer

Read bio

The Modern Application Development Services Landscape, Q3 2024

What Payers Need to Know: Adapting Commercial Offerings Under the New Administration

What Payers Need to Know: Adapting Commercial Offerings Under the New Administration

What Payers Need to Know: Adapting Commercial Offerings Under the New Administration

What Payers Need to Know: Adapting Commercial Offerings Under the New Administration

What Payers Need to Know: Adapting Commercial Offerings Under the New Administration

What Payers Need to Know: Adapting Commercial Offerings Under the New Administration

What Payers Need to Know: Adapting Commercial Offerings Under the New Administration