NovaBridge is a mid-size insurance brokerage managing policies for 12,000+ SMB clients across commercial property, liability, and professional indemnity lines. With 28 operations staff handling everything from policy renewals to claims initiation, they were efficient — by 2018 standards. By 2024, the cracks were showing.
Every routine task required navigating five separate systems: a legacy policy management system, a Salesforce CRM, a third-party claims portal, SharePoint for documents, and an Excel master tracker that had somehow become the connective tissue holding everything together. Checking a single client's policy status required logging into three systems, cross-referencing two others, and often a phone call to a colleague.
"Our ops team are highly experienced insurance professionals. They were spending 40% of their day on system navigation and data entry — not on the judgment calls that actually require their expertise. We needed technology that worked the way they think, not the other way around."
— Head of Operations, NovaBridge
One interface. Natural language. Every backend connected.
The solution wasn't to replace NovaBridge's existing systems — they were deeply embedded and staff knew them well. The solution was to put an intelligent layer on top of all of them that staff could talk to in plain English. Instead of navigating five tools, they'd type a single prompt. The AI agent would figure out which systems to query, retrieve the data, take any required actions, and respond with a clear answer.
We built a multi-tool AI agent using a function-calling architecture. The agent — powered by Claude — was given 8 tools, each corresponding to a real backend API or integration: get_policy_status(), get_client_summary(), get_open_claims(), search_documents(), send_renewal_reminder(), create_claims_note(), generate_policy_summary(), flag_at_risk_policy(). When staff type a prompt, the agent reasons about intent, selects the right tools, calls them in sequence, and synthesises a natural language response.
The hard integration problems
The legacy policy management system exposed a SOAP interface from 2007. We built a REST adapter layer that translated modern JSON requests into SOAP calls and normalised the responses. The claims portal had no API at all — we built a structured scraping layer with a daily sync to a local read-only cache, keeping response times under 1 second.
Infrastructure: GCP Cloud Run (serverless, auto-scaling), Redis for conversation context caching, full audit log of every agent action, role-based access so the agent respects user permissions, SOC 2-aligned logging and data handling. Frontend: React + Next.js chat interface with token-by-token streaming and action confirmation UI for write operations.
The ops team adopted it in 2 days.
We ran a 2-week internal pilot with 6 ops staff before full rollout. Adoption was immediate — not because we forced it, but because the alternative (going back to 5 systems) was visibly worse. By week 2 of the pilot, staff were finding use cases we hadn't anticipated: querying multiple clients at once, asking comparative questions across portfolios, drafting client emails directly from policy data.
"I was sceptical that our team would adopt it — insurance ops people are creatures of habit. Within two days they were asking why we hadn't done this sooner. The queries they're running now, in plain English, would have taken 15 minutes each before. The ROI was visible inside the first week."
— Head of Operations, NovaBridge
