Business Context: The Underwriting Bottleneck
In most commercial and specialty insurance lines, submission volumes significantly outpace the number of qualified underwriters. A typical underwriter in a Lloyd's-adjacent team handles dozens of submissions per week, each requiring document review, risk narrative production, pricing justification, and referral decisions. The administrative burden of extracting data from broker submissions, PDFs, spreadsheets, and emails consumes hours that should be spent on risk assessment. AI changes this ratio. An AI decision support tool ingests a submission package, extracts structured data from unstructured documents, flags the risk characteristics that match or deviate from the underwriter's book appetite, and drafts a preliminary risk narrative. The underwriter reviews, edits, and approves rather than building from scratch. In practice, this reduces time-per-submission by 40-60% on the administrative layer while keeping the underwriter accountable for the final decision. FCA regulation in the UK requires that underwriting decisions remain with qualified humans. The AI system is explicitly positioned as decision support, with human approval required before any policy terms are issued. This is the correct posture both for compliance and for practical reasons: AI models can miss context that an experienced underwriter would catch.
Architecture: From Submission Ingestion to Underwriter Review
The system is built around a document ingestion pipeline and a structured review workflow. Submissions arrive as PDFs, spreadsheets, or email attachments. The ingestion layer extracts text using AWS Textract for structured forms and PyPDF for narrative documents. Claude then processes the extracted text against a structured prompt that identifies key risk characteristics: business description, sum insured, geographic exposure, prior loss history, and any coverage requirements specified by the broker. The extracted data is written to a PostgreSQL database with a clear schema that maps to the underwriter's existing risk assessment categories. The review interface is a Next.js frontend that presents the extracted data alongside the source document, highlights where extraction confidence is low, and provides a draft risk narrative generated by Claude. The underwriter can edit the narrative inline, adjust extracted values, and flag sections for colleague review. Once approved, the structured data and narrative are exportable to the firm's policy management system via API. All actions are logged with timestamps and user IDs, creating the audit trail that FCA-regulated firms require under their record-keeping obligations.
AI Components: Extraction, Narrative Generation, and Appetite Matching
Claude handles three distinct AI tasks in this system. Structured extraction: given a broker submission document, produce a JSON object containing all relevant risk fields. Narrative generation: given the extracted fields and the firm's risk appetite guidelines, produce a draft underwriting rationale in the house style. Appetite matching: compare extracted risk characteristics against portfolio guidelines and flag deviations that require referral or additional information from the broker. The system prompt for each task is built collaboratively with the underwriting team during the SpeedMVPs build process. Getting the extraction schema right, understanding which risk signals trigger referral, and matching the narrative style to what the team actually produces are all domain-specific configuration tasks that require underwriter input. This is not a generic document extraction tool. It is calibrated to the specific risk classes and appetite of the team using it. The EU AI Act classifies AI systems used in insurance underwriting decisions as potentially high-risk depending on their scope, which means proper documentation of the system's purpose, limitations, and human oversight mechanisms is required. SpeedMVPs builds this documentation into the delivery package.
Challenges: Submission Quality and Confidence Handling
Insurance submissions are among the least standardised documents in financial services. A small broker sending a hand-completed PDF form is structurally incompatible with a large broker sending a ACORD-compliant XML submission. The extraction layer must handle both, and handle them gracefully when information is missing or ambiguous. SpeedMVPs addresses this by building explicit confidence scoring into the extraction pipeline: each extracted field carries a confidence level (high, medium, low, not found) based on whether the information was stated directly, inferred from context, or absent. Low-confidence fields are highlighted in the review UI so underwriters know exactly where to focus their attention. The second challenge is hallucination risk. Claude, like all LLMs, can generate plausible-sounding risk narratives that contain factual errors derived from its training data rather than from the submission document. The system prompt strictly instructs Claude to base all claims on the provided document and to flag when information is absent rather than invent values. Underwriter review is the final safeguard. The system is not designed to operate autonomously.
Outcomes: What Underwriting Teams Measure After Deployment
The primary metric is time-per-submission. Teams that deploy this tool typically see the document processing and narrative drafting phase drop from 45-90 minutes per submission to 10-20 minutes, with the underwriter spending the remaining time on review, adjustment, and decision. Secondary metrics include submission referral rate accuracy (whether the system correctly flags risks that genuinely required referral) and narrative quality scores from the underwriting team. Longer term, the structured data extracted by the system creates a dataset that did not previously exist: a consistent, queryable record of every submission's risk characteristics, which enables portfolio analysis that was previously impossible without manual data entry. One insurtech team SpeedMVPs worked with used this dataset to identify that a specific industry sector was generating disproportionate claims relative to premium in their book, a pattern that was invisible in their policy management system but immediately apparent in the AI-extracted submission data.
Lessons: Build for Underwriter Trust, Not Automation
The framing of this product matters enormously for adoption. Underwriting teams that perceive AI as a replacement for their judgment will resist it. Teams that perceive it as a research assistant that handles the document processing so they can spend more time on the interesting parts of the job adopt it quickly. SpeedMVPs recommends positioning the tool as an underwriter copilot in all internal communications, with explicit messaging that the underwriter's decision is final and the AI is doing the legwork. The second lesson is to start with a single risk class, not the entire book. Every risk class has different submission formats, different risk characteristics, and different appetite guidelines. Starting with the class that has the highest submission volume and the most standardised document formats gives you the fastest calibration cycle and the clearest ROI signal. Expand to additional classes once the first one is working well.