AI Customer Service Automation for Cor***** Financial, USACase Study

The service team spent much of its day on calls that required no judgement at all: statement requests, document status checks, contact detail updates and appointment scheduling. Those routine contacts interleaved throughout the day with genuinely complex client conversations, which then waited in the same queue. An AI assistant was the obvious answer in principle, but a general-purpose language model answering clients freely represented a serious liability in a regulated wealth management setting.

The compliance constraints were absolute. The system could never produce anything resembling financial advice, a suitability opinion or a performance projection, and every interaction had to be fully reconstructable for review. Salesforce had to remain the single source of truth for client interactions, ruling out any tool that created a parallel record.

Channel coverage mattered too. Because many clients default to the telephone, a text-only assistant would miss the very callers generating most of the load. The vendor products the firm had evaluated were either too rigid to be useful or too open-ended to pass its compliance function. The firm therefore needed a bespoke build in which control, auditability and safe escalation were designed in from the outset rather than configured as afterthoughts.

PapaSiddhi designed the assistant around narrow scope rather than broad capability. An LLM API handles interpretation and phrasing, but every answer is grounded in retrieval from a compliance-approved content set; the model cannot answer from general knowledge. Out-of-scope intents route to a human, and a dedicated refusal and escalation path covers anything touching advice, projections or suitability. That guardrail behaviour is verified by an automated test suite rather than trusted.

The platform runs on Node.js with PostgreSQL holding full interaction logs, including the retrieved sources and the model's response for every exchange. Those records are written to the relevant client in Salesforce through the Salesforce API, so the service team sees a single interaction history, and a compliance-ready audit export was built with SEC and FINRA record-keeping expectations in mind. Twilio Voice and SMS give telephone callers access to the same assistant through speech, while a React console offers agents a live view of conversations and one-click takeover.

Launch followed a supervised pilot in which every response was reviewed before sending, surfacing several phrasing issues before unattended operation began. The approved content set is versioned and owned by the compliance team, so changing what the assistant can say is a review-and-publish action rather than a code release. Anything outside that approved set is unanswerable by design, which gives the compliance function a clear, inspectable boundary around the assistant's behaviour.