Underwriting is the beating heart of the insurance business. But for a company that has evolved into one of the largest specialty insurers in the world, how it assesses risk across more than 100 products makes that process even more complex. So, while Tokio Marine HCC is committed to optimizing underwriting through intelligent automation, Tamer Assaad would rather they leave the human underwriters right where they are.
Assaad, the company’s vice president of digital transformation and financial systems, oversees an intelligent automation team that grew out of a traditional RPA Center of Excellence. This evolution, he says, reflects a broader shift in what automation can achieve and how it gets there.
Agentic AI, Assaad tells Automation Today, has changed the paradigm guiding automation. Organizations should be asking themselves not if AI can automate an entire job or process, but how it can remove the barriers that keep skilled employees from applying their expertise.
The approach is now being tested in specialty insurance underwriting, where Tokio Marine HCC is preparing to put an agentic AI-based submissions-prioritization system into production. The project uses agents to examine incoming information, conduct research and assess submissions against underwriting criteria before presenting the results to human underwriters who can then focus more on the crucial work of risk assessment and less on the administrative work increasingly associated with the job.
Finding the Work Around the Work
Specialty insurance presents an unusual automation problem. Tokio Marine HCC’s broad offering of specialty products results in many individual business lines with different risk appetites that require varied expertise.
The underwriting project emerged after the automation team began examining submissions intake. Assaad says discussions with underwriters revealed that the most significant issue occurred before the underwriting even began.
“The submission volume you get is asymmetrical to the number of underwriters,” he explains. “There’s no way for those underwriters to look at every single submission.”
Historically, underwriters could face submissions containing varying combinations of emails, applications, loss runs and other documents. They had to research the applicant, examine the material and compare it all against underwriting requirements before determining whether a risk was one the company wanted to consider.
Assaad says an underwriter could spend 20 or 30 minutes reviewing a submission before discovering a characteristic that should have disqualified the applicant before they even started. Meanwhile, potentially attractive opportunities might remain elsewhere in the queue, or worse, be snapped up by a competitor who addressed the customer’s need sooner.
The intelligent automation team developed a pilot system that uses multiple agents to process the incoming material. One set of agents extracts relevant information from structured and unstructured content, while others compare it with underwriting guidelines or perform outside research. The results are used to prioritize submissions according to criteria established by the insurer.
“It doesn’t do anything to replace the expertise or judgment of an underwriter. But what makes it so interesting is it clears the path,” he notes. “It really facilitates them getting to the critical submissions quickly, which has the potential to drive top-line revenue with more profitable bottom-line results.”
The system recommends which submissions deserve attention first rather than deciding whether the insurer should write the business. The underwriter, freed from more time-consuming work, can now review the prioritized submissions and use their specialized knowledge to decide whether to accept the business or not based on the merits of the case.
Reconsidering Design
Designing a successful pilot, as many organizations have found, is relatively easy. Getting an agentic system into production takes considerably more work. To make it happen in this case, Tokio Marine HCC has had to reconsider how it designs automations.
Assaad says one early lesson involved the underwriting guidelines themselves. The team initially supplied the agents with an entire underwriting manual. It subsequently narrowed the material to the portions most relevant to individual assessments, which he says improved both performance and specificity.
The company also involved underwriters directly in development. Assaad says their participation served two purposes: subject-matter experts could identify process nuances the automation team might otherwise miss, and employees who would ultimately use the system had a role in shaping it from the beginning.
Tokio Marine HCC is using UiPath Agent Builder and Maestro in the project, among other technology, and brought in Ashling Partners to work alongside its internal team. Assaad says the outside resources provided additional development capacity as well as another perspective when the team encountered design problems.
Trust on The Path to Scale
As of publication, the underwriting system is in user acceptance testing and is approaching production. Assaad says the company intends to evaluate it against existing business metrics including submission-to-quote, quote-to-bind and submission-to-bind rates, as well as speed-to-quote and the amount of time underwriters spend examining submissions that ultimately are declined.
Explainability has been another consideration and vital in building trust with the underwriting team. Assaad says the system records why it assigns a particular priority or score, allowing underwriters to examine the rationale rather than simply accepting an AI-generated recommendation.
“Underwriters were able to look at the reasoning for a decision,” he says. “Everything was auditable and explainable.”
Security and reliability are also non-negotiable in an industry as highly regulated as insurance. Assaad says Tokio Marine HCC involved its IT security organization and placed particular emphasis on prompt discipline. When information cannot be established, he would rather have the system escalate the issue for human review than manufacture an answer.
“If you can’t find it, there’s higher value in the agent saying ‘I can’t find this, elevate this particular area to human review,’ than to try to fill in the gaps,” Assaad explains.
That philosophy also informs Assaad’s response to concerns about autonomous agents behaving unpredictably, which recently has played out in public for OpenAI and others.
“The answer to agents going rogue isn’t to avoid them altogether or avoid automation, but it’s architecting it in such a way that it can’t take any actions that you’dregret,” he says. “So, depending on what your risk appetite is, narrow that circle.”
Tokio Marine HCC will now be watching what happens as the underwriting application moves beyond testing. Assaad says the company wants to determine not only whether the system affects revenue and policy conversion, but whether prioritization changes the composition of its risk portfolio and whether faster responses affect relationships with brokers.
The longer-term opportunity he sees is applying similar approaches across additional specialty lines, despite differences in their underwriting logic. But his framing of the opportunity keeps the human role intact.
“We didn’t automate the underwriter,” he states. “We identified human expertise as an advantage and automated everything that stood between the underwriter and them doing their best work.”
