Designing AI insurance coverage software to augment commercial underwriting throughput

Applications were waiting, and brokers don’t wait

Element AI partnered with an insurer that wanted AI to help its commercial underwriting team. The team couldn’t process all of its underwriting requests in time. Brokers send the same application to several insurers and the first to answer often wins, so every application in the queue was a business opportunity at risk.

The problem wasn’t skill. Management trusted their underwriters and their decisions, but they couldn’t hire and train new ones fast enough to keep up. They wanted to keep a similar workflow and make no compromise on the quality of decisions.

I joined the insurance solutions team for this project because of my experience turning complex problems and regulatory requirements into clear interfaces. Two constraints shaped everything that followed. The solution had to scale, so that a growing share of applications could be handled fully or partly by AI. It also had to be easy to tune over time as risk appetite or market conditions changed.

Shadowing underwriters changed the brief

Insurance was a new domain for me, so I started by shadowing underwriters at work. I watched how they evaluated an application, what they looked at first and what made them stop and look closer.

Their software was the surprise. It came from large vendors that couldn’t adapt to this insurer’s needs. Simple tasks meant moving between screens and retyping information that should already have been there. Management had described the throughput problem accurately, but none of our conversations with them had mentioned the tools.

That changed what we set out to do. A better workflow would help underwriters even without AI, and it would give the AI a better place to live. One principle also came out of these early conversations: the human stays at the center of every decision. In commercial coverage, a single wrong call can mean very large losses for the insurer.

A workspace of its own, connected to their data

The first idea was to add AI features to the insurer’s existing tools. After the shadowing sessions I argued against it, because those tools were too rigid to build on. We opted for a standalone platform for underwriters, connected to the insurer’s existing application data streams.

uPartner became the underwriters’ marketplace for inbound applications from all brokers. Instead of a plain inbox, the overview showed every application and the step where it had stopped: accepted by the AI model, rejected, or waiting for an underwriter’s review. The AI’s part in the work was visible from the first screen.

This put underwriters’ expertise where the model couldn’t process an application on its own, and left the routine applications to run through automatically.

Every step shows what the AI decided

Each application moves through five steps: submission, segmentation, assignment, risk assessment and coverage. A dark rail on the right shows which steps the AI decided and which ones need an underwriter to verify. When a step needs a person, automatic processing pauses until they confirm or modify it.

Comparing a new application against past ones was one of the slowest parts of the job and could take an underwriter hours. In uPartner, the AI finds similar applications and shows them beside the current one, with the package or risk decision and the date the policy was contracted. The underwriter can adjust coverage with those references in view.

Segmentation. The AI recommends a package and explains its signals, and the step waits for an underwriter to verify it.

At the coverage step, the coverages with low confidence rise to the top under “Requires verification”. The coverages the AI determined on its own sit below, already decided. Management can also change how each step is processed, which gives them control over automatic treatment as their needs change.

Coverage. Three coverages need an underwriter, and the rest arrive already decided.