Blog: Payer-Customized Generative AI Models Deliver Strategic Impact

By John Pyhtila, PhD, Chief Product and Strategy Officer and Gokul Varadharaj, Chief Revenue Officer & Co-Founder

For payers, one of the best ways forward to realizing the potential of generative artificial intelligence is leveraging open-source models for their own use cases based on their proprietary data.

Payers that create custom GenAI solutions by tuning models using their own data in a completely secure environment will gain significant advantages, including:

  • Intellectual property protection
  • Increased accuracy for their populations and network
  • Explainability of results
  • Audit trail for compliance and legal needs
  • Mitigated compliance risks

Except perhaps for the largest health insurers, developing large language models (LLMs) from scratch will be too costly and take too long to build, especially given the scarcity of GenAI talent and proven, built-from-scratch LLMs that solve key use cases. Conversely, while both open-source and proprietary off-the-shelf solutions could cut out some manual work and costs, these generic solutions based on questionable data provide payers with no differentiation or strategic benefit and open up potential risks of sensitive data leaking into the public domain.

Payer-customized GenAI delivers exponentially more scale and enterprise value because it leverages all the exciting possibilities of this technological advance but tailored exactly to solve a payer’s specific business needs using the organization’s proprietary set of data, all while reducing risks.

Customize to optimize GenAI

A payer-customized GenAI tech stack, including data, security, infrastructure, models and applications, offers a springboard for competitive advantage in managing medical costs, member experience and outcomes and provider engagement. “Generative AI is unique in its ability to impact the entire value chain and to drive productivity and growth in a way that establishes a new performance frontier,” according to Accenture’s January 2024 report Reinvention in the Age of Generative AI.

Unlike machine learning, which analyzes data sets to find patterns and make predictions, GenAI intakes massive amounts of data, uses algorithms to train and learn patterns, and then creates (generates) new content by synthesizing pertinent data that addresses the specific ask of the user. LLMs deliver content as text while large multi-model models (LMMs) can also incorporate video, audio and images, which is particularly valuable for healthcare.

Payer-customized GenAI models provide far greater competitive and strategic promise than open-source options for three major reasons:

1.  Payers need to keep core business data secure, private and exclusive.

Many payer operations are highly manual and built around large datasets, so are ripe for improvements through LLMs/LMMs. But core operations including underwriting, claims processing, coding, payment integrity and back-office analytics are based on highly confidential data that is proprietary to the payer and must be kept secure. Sharing this data with publicly available models, both open-source and proprietary, is a non-starter.

Training LLMs and LMMs with proprietary data greatly reduces the potential for inaccuracies, known as hallucinations, that can occur with open-source models using public data of varying quality and reliability. Creating more accurate private and customized models also greatly increases internal trust for GenAI, which accelerates adoption and ultimately business impact for the payer organization.

2. Payer-customized GenAI models deliver greater effectiveness.

Beyond keeping data protected, customized GenAI, based on their crown-jewel data, gives payers a market advantage with insights tuned to their populations, policies, procedures, and other circumstances. These models are trained on local nuances within the data, including patient populations, geographic differences, provider partners, internal confidential documents, and more. Results from customized GenAI models also improve over time as new data is added and refinements based on user queries and answers are made.

3. Integrating individual models maximizes performance.

Rather than trying to deploy a monolithic model for all needs, payers will have more success, faster, with customized LLMs/LMMs designed to solve specific business challenges and using data for that business area. While a single model could potentially answer queries involving underwriting, STARS, and medical policy, each area depends on both some shared data and unique data assets. Models for each area would generate much more accurate responses and also provide enhanced ability to control internal access to specific data assets.

These LLMs/LMMs can be integrated with a master model routing queries to the correct LLM or LMM for a response. This mesh of models results in a seamless user experience and enhanced model performance.

Usable data is a customized GenAI cornerstone

Just as with publicly available GenAI, data integrity is central to building and scaling successful private GenAI solutions. Yet 56% of executives across industries cite lack of data readiness as a top challenge in adopting AI, according to June 2023 Accenture research.

While payers have become conditioned to settling for the data they can get, they are running into the limits of this approach. GenAI will only exacerbate the problem of old, siloed, incomplete and inaccurate data.

To fulfill its promise as a catalyst of enterprise reinvention and healthcare reinvention by extension, generative AI needs to be built on a foundation of usable data, a new standard of data that meets six criteria: accurate, complete, timely, relevant, versatile and application and use-case agnostic. For GenAI use cases, usable data inputted into LLMs/LMMs would be tailored to the specific data needed each business purpose. Data, then, is the key to GenAI-driven performance improvement, not the “best” models per se.

For GenAI, usable data goes hand-in-hand with modernized technology and data infrastructure. “Generative AI requires a fundamentally different enterprise architecture in which data is more fluid and unstructured and synthetic data become much more important,” according to Accenture’s Reinvention in the age of generative AI report.

As generative AI places higher demands on data, it’s the opportune time to rethink and rebuild in an enterprise architecture around highly trustworthy, usable data as the rocket fuel for innovation and achieving new levels of business performance. Without the necessary investments now for ensuring a foundation of highly usable data, payers will find themselves falling behind at an accelerated rate as GenAI becomes more prolific.

Contact us at if you’d like to learn more about getting the usable data you need to power GenAI and strategic initiatives.

This blog is first in a series about the use of GenAI by Payers. Look for the second blog, about the four pillars for developing payer-customized LLMs/LMMs, to be released soon.