Case Study
A Deep Research Agent for Underwriting
How Sompo International fast-tracked their deep research agent, slashed underwriting research time and improved risk management.
Sompo International are a multi-national insurer and reinsurer, serving corporate clients globally. The fundamental limit to the quantity of business they can write is the time it takes underwriters to triage submissions and then to research pertinent risks affecting the insured.
By developing a Deep Research Agent, Sompo dramatically reduced the time to search for and collate the risks, leaving underwriters to focus on analysis and risk management and shortening their key metric of “time-to-bind” (the time taken from initial quote to final confirmation of the policy)
The Challenge
Sompo International makes its living by understanding risk better than anyone else. When a new submission lands on their desk, underwriting teams need to know everything: what could go wrong, how likely it is, and what it'll cost when it does.
Here's the rub:
During discussions with the underwriters it emerged that they were burning through half their company research time just trawling the web for information.
Even then, there were persistent worries that something could have been missed.
The other half of their time was spent on the actual analysis and meeting prospects' management teams: in other words, the bits that require deep, human expertise.
The client knew there was an opportunity here: cutting the time spent searching would lead to higher throughput and therefore more business being written.
They came to Veratai with a simple question:
Could a Deep Research Agent help?
How Veratai Helped
We started with a short feasibility study to make sure that the business case was there. Good news: not only was it viable, the economic case was compelling enough to make the client’s digital innovation lab sit up and take notice.
Next came the “baseline test”. We evaluated off-the-shelf agents from all the frontier labs - the big names you'd expect, to see how well they did at researching and triaging insurance-related risks.
They all fell short.
Why?
Two reasons:
There is a great deal of tacit knowledge in the underwriting process - and even frontier LLMs don’t have it
There are product-specific, industry-specific and insured-specific factors that make each research task unique.
Underwriters will naturally hone in on this but the off-the-shelf agents lacked the understanding of this nuance.
To resolve these shortcomings, we embedded ourselves with stakeholders across the business, mapping out how different product areas approached research. The variations were fascinating. This wasn't a uniform process - research tended to be insurance product-specific, but there were enough commonalities for us to determine that a single agentic backbone could be made to work. We distilled our learnings into a research framework that captured both the commonalities and the essential differences. This became our north star.
By codifying this knowledge into a “research framework” and scaffolding an LLM with it, we were able to distil decades of expert experience into a research agent which could find, filter and prepare the right resources for the underwriters.
What we ended up with was a custom AI Agent designed from the ground up to handle the specific complexities of insurance research, paired with an End User Application in which they could explore it.
The Outcome
Six underwriters. Six weeks. Real submissions.
The results?
Data collection time? Slashed.
From the pilot we esimated 2 hours per research task - that’s around 8 hours each week for an underwriter.
Coverage? Comprehensive
The agent found "nearly all" relevant risks that human researchers identified - and this was just the prototype.
User verdict? Unanimous.
Every single pilot user wanted this in production. Yesterday.
We're now building the production system with a full rollout across the business in the works.
The benefit: underwriters will soon be able to spend their time on what matters - weighing risks, meeting with the insureds and adjusting pricing - rather than trawling through news-feeds or social platforms.
This is not just efficiency; this is competitive advantage.