Convex Transforms Specialty Insurance Underwriting with Generative AI

Convex is a specialty insurance company with operations in Bermuda, Europe, the UK, and the US, backed by over $3B in long-term capital and rated excellent by AM Best and Standard & Poor’s. Their leadership sees running the business on data and AI as a top priority.

01 The Challenge

The hundred-page risk reports as the bottleneck in the underwriting workflow

At Convex, before an underwriter could price a risk, a subject-matter expert read a 100-page engineering report – technical, dense, different per broker – then prepared a structured summary. The underwriter waited on the summary. The book grew at the pace of the SME queue.

“The underwriters were never a problem. The bottleneck was not their judgment but the manual work upstream of the judgment.” · Head of Operations · Convex

Specialty insurance underwriting sits in a narrow class of workloads where generative AI can take over manual document processing. The expert read stays with the underwriter. Convex came to Provectus to build a solution that could speed up and scale underwriters’ decision-making by using GenAI to make sense of risk reports, faster.

02 The Approach

Prototype against a tuned LLM. Then make the prototype operational.

The work ran in two phases, each gated on a measurable outcome.

The first phase built a working prototype of the document extraction pipeline: a tuned LLM on Amazon Bedrock, an evaluation harness for the whole pipeline and for each subtask, and several iterations over extraction, segmentation, classification, and summarization. The prototype had to clear a quality bar before the second phase started.

The second phase productionized the prototype. Discovery sessions with Convex covered their KPIs, their existing tech stack, their security posture, and their reviewer workflow. Out of that came an LLM pipeline in production, a reviewer cockpit, CI/CD, versioning, monitoring, and a Retrieval Augmented Generation (RAG) workflow tuned to Convex’s own corpus of reports.

03 The Build

Tuned LLM plus reviewer cockpit plus production ML pipeline

The extraction model is a tuned LLM accessed through Amazon Bedrock. Classification components use dedicated models customized on Convex’s own report data. The pipeline runs extraction, segmentation, classification, and summarization as sequenced stages; each stage cites back to the source page of the original PDF.

Output is machine-readable JSON downstream and a reviewable summary document for the SMEs. Every GenAI-generated claim traces to a source region in the underlying report. Every SME edit is logged.

The reviewer cockpit routes edge cases to the SME – adversarial wording, unusual exposure profiles, judgment calls. Everything else runs on the pipeline.

Deployment is on Convex’s own AWS environment. Observability, cost controls, and kill-switches are built in.

04 The Results

Ten-minute report summaries. A hundred times the book against the same team.

10 min

From a 100-page report to a structured summary

Measured against the SME baseline

The cost of reading a report dropped below the cost of ignoring one. Convex can now price and decide on risks that would previously have sat in a queue. The portfolio grew as the book filled with risks the team no longer had to decline for lack of time.

Underwriter-facing edits stay consequential. Every SME review teaches the classifier so that quality rises with volume.

05 What’s Next

Submission Flow: the baseline Convex tuned, packaged for the next carrier

The Convex engagement is the tuned baseline for Submission Flow, the Provectus blueprint for carriers and MGAs running the same submission-to-triage loop. The next carrier starts from the baseline Convex spent a quarter tuning. The Sprint – one week, fixed fee – measures what the carrier’s own book would do against that baseline before any Integrate phase begins.