The Underwriter's Excel Monster

So I walk into this reinsurance company thinking I'd build some flashy executive dashboards. We deliver a chatbot that can handle multiple file formats that the executives are utilizing to derive insights from. The CIO pulls me aside week two and says, "Forget the dashboards. We have this Excel problem that's driving our executives crazy."

He shows me these massive, multi-tab spreadsheets. The kind that take 30 seconds to open and make your laptop sound like a jet engine.

AI Tools Aren't So Smart

Everyone assumes ChatGPT or Copilot can handle any Excel file, right? Upload spreadsheet, ask question, get answer. Simple.

Except it's not simple with reinsurance data. These aren't clean sales reports. These are complex datasets with nested tables, merged cells, and relationships spanning multiple tabs.

The CIO had test questions he'd been trying to get answered for months:

"Give the IBNR data, Federal Exercise Tax and Amount Held from the 2022 TY Less Excl sheet for all ceded contracts for 4Q2024 in tabular format"

"For all contracts with Property type business segment, give the Net Due ORC/(Ceding Co) for quarter to date from the 2022 TY Account Statement sheet, comparing values from both files in tabular format."

The AI Tool Failures

ChatGPT would confidently pull data from completely wrong tables. Mix up "2022 TY Less Excl" with "2022 TY Summary" sheets. Sometimes just make up numbers entirely.

Copilot found the right data but did the math wrong. Would locate correct columns but apply wrong filters or lose track of which file it was working with.

Gemini gave detailed explanations of what it would do, then completely failed to execute. Would hallucinate data or only use one file when you asked for comparisons.

The worst part? They presented wrong answers with complete confidence. No hedging, just "Here's your data" followed by fabricated numbers.

Building Something That Actually Works

We had to build our own solution. The breakthrough: instead of making AI understand Excel directly, create a translation layer between natural language and structured data operations.

Step 1: Smart Excel Ingestion

Standard pandas functions choke on these files. We built smart ingestion:

Key insight: reinsurance Excel files are documents with implicit structure, not just data dumps.

Step 2: Intelligence Index

Instead of storing raw data, we built an index capturing context:

CREATE intelligence_index:
  - Map business terms to actual columns
  - Track relationships between sheets
  - Store sample data for validation
  - Link calculated fields to dependencies
  
EXAMPLE mappings:
  "IBNR" → ["IBNR", "incurred but not reported", "ibnr data"]
  "Net Due ORC" → ["net due orc", "net due orc/(ceding co)", "orc due"]

This maps reinsurance terminology to actual data locations.

Step 3: Query Translation

The magic happens here. Systematic translation of executive questions:

"Give IBNR data from 2022 TY Less Excl for Q4 2024" becomes: locate specific sheet → find IBNR columns → apply Q4 2024 filter → extract → format as table.

Step 4: Multi-File Comparisons

Where most AI tools completely fail:

The Results

Dramatic transformation. The CIO's impossible test questions now worked consistently. Executives got answers in minutes instead of hours.

Bonus discovery: The system found data inconsistencies nobody noticed before. When the same contract appeared with different values across reports, it flagged discrepancies.

Follow-up magic: "Now show me the same data but only for contracts above $50M." The system understood context and built on previous queries.

Security solved: Everything stayed on-premises. No compliance issues with sensitive files.

Technical Lessons

1. Domain expertise > technical sophistication
   Understanding "2022 TY Less Excl" matters more than fancy AI models.

2. Data prep is the real magic
   LLM is just the interface. Intelligence is in understanding reinsurance data structure.

3. Reliability > cleverness
   Better to correctly answer 95% of common questions than fail unpredictably on edge cases.

4. Integration wins
   Success came from generating reports in existing formats, not forcing new tools.

Want to See It in Action?

The best way to understand what we built is seeing it work with real reinsurance complexity. We've set up a demo with sample files that capture the structural complexity of actual treaty accounting reports, cedant analysis files, and loss development triangles.

Try those exact queries that stumped ChatGPT and Copilot. See how it handles tricky multi-sheet queries, cross-file comparisons, and reinsurance-specific business logic.

Ready to see AI that finally gets reinsurance data right? Let's walk through it together with our sample files. Way more fun than it sounds, and you might be surprised what becomes possible when your AI tools actually understand your business.

Try our sample files and see the difference →