The workflow before
The Tungsten Marketplace lists hundreds of automation tools for insurers, from claims intake to underwriting and compliance. Users searched by keyword, which missed context, and anything deeper meant writing queries or asking support. Finding a relevant listing took 15 to 30 minutes, about 40% of visitors left without finding one, and support handled around 150 'how do I find' tickets a month.
What we built
We built a copilot inside the Tungsten Marketplace that answers questions in plain English, such as 'compare document processing tools for underwriting', and returns ranked listings with the reasons behind them.
Under the hood, BERT-based models classify intent and extract entities, a Neo4j knowledge graph maps listings, vendors, use cases and integrations, and a vector index handles semantic search. Recommendation models personalize results by role, insurance line and company size.
The copilot also surfaces new listings and integrations to users who will care about them, and guides them through comparing and evaluating tools. It went live after a 35-week build.
Before and after
| Measure | Before | After |
|---|---|---|
| Time to find a relevant listing | 15 to 30 minutes | 2 to 5 minutes |
| Searches that lead to an evaluation | 5 to 8% | 18 to 25% |
| Support tickets per month | 150 | 40 |
| User satisfaction | 3.2 out of 5 | 4.5 out of 5 |
What's next
The next step is Level 3 (Workflow Agents): having the copilot carry a buyer from shortlist through integration checks and a trial request, with Tungsten's team reviewing the exceptions.
