Context Over Prompts: The Critical Shift for AI Success in Insurance

By: | September 29, 2025

Sean G. Eldridge is the Co-founder and CEO of Gain Life, a venture-backed insurance technology company that helps individuals and organizations return to health, work, productivity, and financial wellbeing. Prior to Gain Life, Sean led a private equity-backed roll-up in the disaster restoration space and held leadership roles at Johnson & Johnson, Procter & Gamble, and Weight Watchers, launching new products and services that leverage the power of behavioral science and technology. Sean earned his B.S. in Management Information Systems from Rochester Institute of Technology and MBA from Harvard Business School. He resides in Cambridge, MA with his wife, son, and corgi.

A few years ago, I attended an insurance technology conference where one session after another focused on the art of the perfect prompt. We all scribbled down example instructions, as though capturing secret recipes, convinced that if we found the right wording, artificial intelligence (AI) would transform our businesses. It felt a bit like learning to speak a new language overnight. Fast forward to today, and the conversation has shifted completely.

The magic isn’t in the phrasing anymore. It’s in the information the AI has at its disposal.

As these models have grown smarter, they no longer require detailed instructions. Instead, what matters now is ensuring the AI has access to the right information: what has worked for your business, which data drives actual outcomes, and what’s relevant for the problem at hand. This gave rise to what’s now called context engineering, the practice of intelligently managing what information the AI can access when making decisions.

The Shift in Practice

In 2023, evaluating a workers’ compensation claim with AI meant writing a prompt that specified every step. First, check the injury date. Verify employment status. Review medical documentation. Assess state regulations. Calculate benefit rates. Insurance companies maintained libraries of these prompts, some running thousands of words, trying to capture every possible scenario. The use cases were limited to simple Q&A and basic document generation.

Today’s models already understand what claims processing involves. They know medical review means checking diagnoses against treatment protocols. They understand that employment verification requires confirming active status at the time of injury. The intelligence is there.

But intelligence without information is useless. A model might understand claims processing perfectly, but if it cannot access the injured worker’s previous treatment history, the employer’s policy details, or the relevant state regulations, it cannot make decisions. The challenge shifted from teaching the AI what to do to ensuring it has the right information to work with.

This shift unlocked entirely new possibilities. Instead of just answering questions about coverage, AI is now handling complex multi-step processes: full claim adjudication, fraud pattern detection across multiple cases, real-time treatment authorization, even predicting claim trajectories based on historical patterns. With the right context, AI can handle everything from customer inquiries to underwriting decisions to fraud detection.

3 Challenges of Context Engineering

Making context engineering work in insurance isn’t straightforward. The theory is clear, but implementation requires mastering three distinct challenges:

1) Context Selection: Knowing what information to provide has become the new critical skill. Even as context windows grow from 4K to 100K-plus tokens (roughly 3,000 to 75,000 words of text), more isn’t always better. Do you provide full medical history or just recent treatments? Raw doctor’s notes or summarized findings?

The expanding context windows mean you can include more, but that makes intelligent selection even more important. Success requires both technical understanding of how AI processes information and insurance knowledge about what actually drives claim outcomes.

2) Security Architecture: AI needs information in its context to function, creating new vulnerabilities. Users might try to extract information about other claimants or reveal proprietary patterns through prompt injection, attacks that embed hidden commands in normal requests.

The solution requires dynamic context assembly based on who’s asking, filters preventing inappropriate data from entering context, and audit trails tracking what context was assembled for each interaction.

3) Integration Reality: Insurance data is scattered across countless systems, with claims in one place, policies in another, and medical records arriving by fax. A workers’ compensation claim might need data from HRIS systems, multiple EMRs, case management platforms, and regulatory databases.

Systems must also remember previous interactions rather than starting fresh each time. Doing this successfully, requires unified data access layers, semantic reconciliation, and real-time pipelines that assemble context on demand.

Looking Forward

Moving from prompt to context engineering means rethinking how insurance companies deploy AI. The competitive advantage won’t come from having the smartest AI, but from giving that AI access to the right information at the right time, while maintaining security and compliance.

Ask questions. Can you explain why context needs change between first notice of loss, initial reserve setting, and claim settlement? How will you handle role-based context assembly for different users? Can you pull real-time data from legacy policy administration systems? What happens when a claim spans multiple states with different regulations? How do you prevent prompt injection from exposing other claimants’ data? Can you maintain context across a claim that takes weeks to resolve?

Use these questions when evaluating vendors. If a vendor can’t explain their integration architecture or security model, they’re not the right partner. If they talk about AI capabilities but not about data access, they haven’t solved the real problem. The best vendors understand that success in insurance AI depends less on having the most advanced model and more on understanding the complexity of insurance data, the reality of legacy systems, and the critical importance of security and compliance.

The winners in insurance AI won’t be those with the cleverest prompts. They’ll be the ones who master providing the right information while maintaining the standards that insurance demands. &

More from Risk & Insurance