Within my current role at Continental Heritage, I saw a repeated manual bottleneck in commercial underwriting, and no enterprise AI foundation to build on. Nobody asked me to fix either one. I shipped a bond document analyzer that reads bond and financial documents and auto-fills bond paperwork with human-in-the-loop review, and I personally ran the privacy, data, and security due diligence that brought Claude Enterprise into the company and got it fully deployed. That combination is now the foundation for a broader AI copilot vision for underwriting that I'm actively pursuing.
Nobody asked me to build this. I noticed the same manual bond-filling and system-coding work resurfacing every time a new bond document showed up, and I noticed the company had no vetted enterprise AI platform and no data-governance path that any team could safely build on. Both problems sat within reach of my role, so I treated them as product problems, not IT problems, and took ownership of solving them myself, on top of my existing scope.
I started with the narrower, more concrete piece - the bond analyzer - because it was a real, bounded workflow I could ship and prove value on quickly. I built it with the longer-term opportunity in mind: if I could show a well-scoped AI tool could safely take real manual work off underwriters' plates, I'd have the credibility and the infrastructure to propose something bigger.
The bond document analyzer reads a bond document alongside its principal's financial information sheet, and produces a structured analysis of what the bond actually requires - the specific fields, figures, and conditions an underwriter needs to act on, instead of reading the whole document cold every time.
It fills the bond document itself, with a human-in-the-loop click to confirm before anything is finalized, and it scales - the same tool can take financial data for 100+ insurers submitted at once and auto-fill the corresponding bond paperwork for each one. That directly eliminated a repeated manual process: instead of re-keying data or writing new system logic every time a new bond format showed up, the tool now handles extraction and fill, and a human still signs off.
The bond analyzer needed a capable AI platform behind it, and the company didn't have one yet. Rather than build on an ungoverned or unapproved tool, I ran the actual enterprise onboarding process myself: reviewing Claude Enterprise's privacy and data-processing terms, defining retention controls, confirming audit and access controls, and setting up the Microsoft 365 connector so it worked inside our existing environment instead of alongside it.
That due-diligence work is what let Claude Enterprise move from an idea to a fully deployed platform the company now uses in production. It also means the bond analyzer, and anything I build after it, sits on governed, reviewed infrastructure instead of a one-off integration nobody else can safely extend.
With no existing enterprise AI foundation, I weighed three ways to get real AI capability into production.
| Option | Governance | Time to Value | Extensibility | Verdict |
|---|---|---|---|---|
| Build a fully custom in-house model/tooling stack | High effort to govern well | Very slow | High, but reinvents fundamentals | Rejected, too slow to prove value |
| Let individual teams adopt disconnected AI tools ad hoc | Low, inconsistent | Fast per team | Low, no shared foundation | Rejected, creates ungoverned risk |
| Run enterprise due diligence, onboard one governed platform | High, reviewed & controlled | Moderate | High, shared foundation for future AI work | Selected |
Running the due-diligence process myself took longer than just asking a team to try a tool, but it's the reason Claude Enterprise is now used broadly and safely across the company, and the reason the bond analyzer has real infrastructure under it instead of a fragile one-off integration.
Extend the same grounded, human-in-the-loop approach from bond documents to the wider underwriting workflow - submissions, financial statements, supporting documentation.
Automate extraction and calculation of underwriting-relevant financial ratios and signals, the same way the bond analyzer already automates bond-specific data.
Structured, source-backed case summaries, and automatic surfacing of contradictions or missing information before an underwriter starts reviewing.
Suggested follow-up questions or missing-document requests based on what the copilot can already see is incomplete.
My evaluation philosophy: a faster answer isn't a better answer unless it's also reliable, reviewable, and adopted. I'd evaluate Phase 2 across speed, quality, trust, and adoption together, not productivity alone - the same discipline that shaped how I scoped and governed Phase 1.
Ownership without being asked. I didn't wait for a mandate to fix a workflow problem or bring AI into the company. I saw both, and did the unglamorous governance work most people skip.
Vision means more when it's paired with execution. The broader copilot roadmap is credible precisely because it's coming from someone who has already shipped and governed a real piece of it, not just described one.
Infrastructure work is product work. Due diligence on data-processing terms, retention, and access controls isn't visible in a demo, but it's what let Claude Enterprise actually go into production instead of staying a pilot.
Scope discipline compounds. Shipping the narrow bond analyzer first, instead of pitching the full copilot on day one, is what made Phase 2 a credible next step instead of a speculative pitch.
