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

The client's service team was fielding a high volume of calls that required no judgement — statement requests, document status, contact detail updates, appointment scheduling — while genuinely complex client conversations queued behind them. An assistant was an obvious fit, except that in regulated financial services a general-purpose language model answering client questions freely is a serious liability. The system could not under any circumstances produce anything resembling financial advice or a performance projection, and every interaction had to be reconstructable for compliance review. Salesforce was the system of record and the team was not willing to have a second source of truth for client interactions. There was also a practical issue with channel: a meaningful proportion of their client base defaults to the phone rather than chat, so a text-only solution would miss the callers generating most of the load. Existing vendor products in this space were either too rigid to be useful or too open to be approved by the client's compliance function.

We built the assistant with deliberately narrow scope rather than broad capability. The language model handles interpretation and phrasing, but answers are grounded in retrieval from a compliance-approved content set — it cannot answer from general knowledge, and out-of-scope intents route to a human rather than being attempted. A refusal and escalation path covers anything touching advice, projections or suitability, and that behaviour is tested as a suite rather than trusted. Every exchange is logged in full, including retrieved sources and the model's response, and written to Salesforce against the client record through the Salesforce API so the service team sees one interaction history. Twilio provides both voice and SMS channels, so telephone callers reach the same assistant through speech rather than being pushed to chat. The React console gives agents live visibility of assistant conversations and one-click takeover. We ran a supervised pilot where every response was reviewed before the assistant handled live traffic unattended — a slower path to launch that the compliance team required and that surfaced several phrasing issues we would otherwise have shipped. Content in the approved set is versioned and owned by the compliance team rather than by engineering, so updating what the assistant can say is a review-and-publish action rather than a code release. Anything not in that set is simply unanswerable by design. Related: AI & ML development and our fintech practice.