Bringing Smart Underwriting to Health Insurance

Contributors: Stacy Dubovik, ScienceSoft’s Financial Technology and Blockchain Researcher, and Alex Savanovich, ScienceSoft’s Senior Data Scientist.

Here is what I noticed in a few recent studies covering the health insurance underwriting sector: insurers that invest in AI-powered automation report a 35–40% reduction in operational costs, a 20–30% increase in revenue, and a 3x spike in customer satisfaction rates – all driven by more efficient underwriting processes. It’s no wonder that seeing this trend, 72% of health payers pursue aggressive digital transformation strategies, and 65% plan $10M+ investments in cognitive automation by 2025. Most of my health insurance clients, regardless of their digital maturity levels, share this sentiment and put a big focus on enhancing their underwriting technology.

Then, why have intelligent tools for automated health underwriting not been widely adopted yet? It turns out, two thirds of health payers struggle to accurately plan their underwriting AI initiatives. Insurers who turn to me for advice mention vague cost and value estimates for emerging AI techs as the major planning hurdles. Although the costs of implementing AI are democratizing gradually, some payers fear that launching a standalone tool like a white-label chatbot may incur substantial extra investments in new data and integration solutions and, as such, hamper the ROI. Plus, the data privacy and compliance concerns often restrain AI launch. In my practice, there were cases when the hesitation about potential HIPAA violations pushed payers to abandon their AI projects midway.

The good news for health insurers is that there are definitely ways to bring smart automation into underwriting without breaking the bank or compromising PHI security and legal adherence. To give you an idea of the investments, in ScienceSoft’s recent projects, we managed to deliver standalone compliant solutions within the budget of $100,000–$350,000.

Below, I share the options that can become a win-win starting point for your underwriting transformation journey.

Specialized AI Tools for Health Underwriting and How to Safely Integrate Them Into Your Workflows

Smart copilots for 2x higher underwriter productivity

Since the release of ChatGPT, many health insurance clients I met have been seeking ways to incorporate the same natural language assistance into its workflows. The success story of Markel, which pioneered smart underwriter copilots in specialty lines, showed impressive outcomes: a 100% increase in underwriter productivity. Decision-ready risk data was ready within minutes due to real-time access to the required proprietary and public data, automated risk document review, and guided health risk scoring.

Imagine your underwriter assessing a customer with pre-existing or ongoing health issues. Traditionally, complex cases require consultations with medical professionals to evaluate condition severity and associated risks. Now, GenAI assistants can instantly auto-interpret specific medical conditions and their implications for coverage, minimizing the strain on your underwriters and allowing them to produce quotes faster.

Where can you get a similar tool, though? Using consumer chatbots such as ChatGPT is out of the question due to the lack of specific model knowledge, as well as data privacy and compliance implications. For now, I see custom implementations based on market-available models as the only feasible and safe way to bring intelligent assistants to the health underwriting loop. My colleagues from ScienceSoft’s data science department usually employ prebuilt GenAI models (GPT-4, LLaMA, Claude, etc.) and train them on the required health insurance specifics and compliance rules using cost-effective retrieval augmented generation (RAG) and prompt engineering techniques. You can further reduce costs by adding copilot interfaces to existing underwriter apps or in-browser extensions instead of building standalone AI applications.

LLM solutions for a 10x reduction in manual data processing efforts

Large language models (LLM) can capture, consolidate, classify, and summarize risk data from heterogeneous health insurance documents, such as applications, health statement forms, lab test results, and prescriptions. This eliminates up to 90% of the most tedious underwriting routines, accelerating the time to quote by up to 12x.

Off-the-shelf tools like Cytora and Expert.ai are go-to choices for automating pre-enrollment workflows, but it is costly to customize them to a broader scope of tasks. On the other hand, custom LLM solutions will allow you to reuse model logic across other medical insurance functions (e.g., in underwriting and claims). You don’t need to build your own models from scratch. Applying retrieval augmented generation (RAG) and minimal parameter tunes to commercial LLMs is usually enough to obtain an accurate solution.

Regardless of the nature of the GenAI model, consider deploying your solution either on-premises or in a private cloud. This is the safest option in terms of protecting PHI processed through intelligent models and maintaining compliance with HIPAA and similar regulations.

Intelligent decision-making engines to process 90%+ of applications within munites

Experts from McKinsey believe that advanced machine learning models can potentially automate almost the entire underwriting flow, cutting the cycle to just a few seconds. Looking back at my previous projects, it’s unlikely one can build a truly end-to-end automated process, as non-trivial health cases will still require a human touch. But for standard health insurance applications, straight-through processing is a reality.

Intelligent decisioning solutions can auto-process applicant data, profile an individual’s or a group’s overall health risks, and immediately calculate tailored quotes. Best-in-class proprietary engines handle over 90% of standard applications outright, and the degree of automation can be higher than 95% if you mainly deal with low-risk cases. This opens immense opportunities for freeing your underwriters’ capacity and driving insureds’ satisfaction.

Compliance above all: how to get explainable AI for insurance

One common concern among my health insurance clients is the compliance of AI decisions with federal insurance plan terms (ACA, Medicare, Medicaid, etc.), FTC guidelines for ethical AI/ML use, NAIC regulations, internal policies (e.g., for short-term plans), and more. Given AI’s ‘black box’ decision-making logic, many worry it may be hard to control the accuracy of auto-writing and prove adherence to regulations.

The solution is creating explainable AI models that provide a clear rationale behind their outputs. My teammates from the data science department usually apply techniques like LIME and SHAP to interpret AI logic and incorporate requirements for source documentation in prompt templates to trace AI’s reference points. For example, when deciding on applicant enrollment in an ACA health plan, the model would reference PPACA provisions, CMS guidelines, EHB requirements, NAIC rules, as well as your corporate risk adjustment and pricing policies. Luckily, the big technology players are aware of regulatory scrutiny, so major AI platforms like Azure Machine Learning, Amazon SageMaker, and Google’s Vertex AI come with built-in explainability toolkits.

Not Ready to Automate With AI? Rule-Based Tools Are Still Effective

Rule-based automation still prevails in the underwriting domain, so for many payers, the near-term objective is to make the most of their previous-gen algorithms. If you want to replace your conventional automation tool with something better but still non-AI, low- and no-code platforms like Microsoft Power Apps, Pega, and Appian are your way to quickly launch dynamic and resilient rule-based underwriting solutions. The intuitive drag-and-drop interfaces of these platforms make it easy for underwriters to design and modify automation rules without IT staff involvement. As far as the cost goes, building a health underwriting automation solution on a low-code platform may be up to 70% cheaper compared to development from scratch.

Pragmatic Solutions for Zero-Touch Risk Data Exchange

Integrations are not just an efficiency driver but a necessity for precise risk assessments. Corporate and third-party risk databases, agent and customer apps, policy management tools, claims solutions — all these need connections to your underwriting system for seamless data sharing. Plus, some dynamic data from emerging sources, like connected medical devices for real-time patient monitoring, comes in formats and volumes that are not manually processable by nature.

Whether you plan to engineer a new underwriting solution or upgrade your aging system, prioritize service-oriented, API-first architectures and cloud deployments. This will ensure the solution’s interoperability with external data sources and will let you easily establish new integrations in the way ahead. Healthcare providers and public services that operate digital PHI usually provide ready-to-use FHIR APIs for secure access to medical data, so you only need to build client APIs to automatically obtain data to your system.

What if your heritage systems do not support API-enabled integrations? One way is to build dedicated custom connectors, but the option doesn’t come cheap and complicates integration maintenance if you retain multiple legacy tools. A cost-effective workaround in this case would be implementing API-driven integration middleware.

Blockchain EDI may be a secure emerging alternative to API-enabled HIE solutions. However, rolling out a full-scale private network is only feasible for health insurers with numerous medical partnerships and extensive digital data flows. The option enables fully automated multi-party sharing of patient health, risk, reinsurance, and eligibility data using smart contracts, which contributes to increased underwriting efficiency.

Laying the Basis for Controllable Underwriting Improvements

The benefits of enhanced underwriting are real — but you need a robust KPI system to measure your gains. From my experience, metrics like underwriting cycle time, quote-to-bind, cost per quote/bind, loss ratio, retention rate, and CSAT may give abundant insights into the actual value of digital health underwriting initiatives. Consistent growth across these metrics signals positive results attributable to improved underwriting specifically and indicates the initiative’s success.

If you need advice on the right digital tools, implementation options, or KPIs for your specific case, don’t hesitate to contact me or other consultants at ScienceSoft.