AI Workflows for Faster Customer Inquiry Resolution
How we would approach inquiry routing and response drafting across siloed departments in a regulated firm.
A financial services firm routes customer inquiries by hand across banking, investments and insurance teams, each with its own systems and knowledge base. This scenario walks through a human-in-the-loop routing and drafting system built for an auditable environment.
What we would agree up front
Measured first-time routing accuracy against a labelled baseline
Every drafted response cites its source policy
Complete audit trail for each automated classification and draft
A composite scenario showing how we approach this kind of problem, not a specific client engagement.
The situation
Inquiries are often sent to the wrong team first, agents spend much of their time searching Confluence, SharePoint and PDF policies, and regulatory response deadlines for some inquiry types are at risk.
Constraints we would design around
Customer data must stay within the firm's environment. Backend systems are old, without modern APIs. Every automated step needs an audit trail a compliance reviewer can follow.
Approach
Map the inquiry lifecycle and failure points first. Then automate classification and routing, retrieve relevant policy for the agent, and draft a response with citations, leaving the final decision with the agent. Full autonomy is not the goal; fewer handoffs and faster, better-grounded answers are.
Architecture
Agent workspace in React with a knowledge panel and response editor. FastAPI services for classification, retrieval and drafting. A RAG index over internal knowledge with permission-aware retrieval, so agents only see sources they are entitled to. Models served through the firm's own cloud tenancy, and every action logged for audit.
Design notes
Citations — linking every drafted sentence back to the policy it came from — are what make compliance teams comfortable, and agents end up using them to learn. Routing accuracy should be measured against a labelled sample before and after launch, not assumed.