A non-bank mortgage lenderFinancial services, non-conforming mortgage lending
Dual-Audience Mortgage Agent
One agent that recognises whether it is talking to a broker or a consumer, matches slang and shorthand to the real product catalogue, and stays inside strict compliance boundaries on every reply.
Before
2 audiencesBrokers and consumers on the same channels, with generic chatbots mismatching shorthand such as refi or no money down to the wrong loan type
After
1 agentClassifies the enquiry on arrival, adjusts depth and tone, matches the catalogue and answers safely every time
Overview
Broker depth. Consumer plain English. One agent.
A non-bank mortgage lender was handling a high volume of enquiries from two very different audiences on the same channels. Brokers wanted quick, technical answers about product eligibility, documentation and edge cases. Consumers wanted plain-English guidance about what they might qualify for.
Every response had to stay inside strict compliance boundaries, with no specific quotes, no financial advice and no approval promises, across a product catalogue large enough that generic chatbots routinely matched the wrong loan type and cost leads on both sides.
The problem
Two vocabularies, one catalogue, zero room for error.
Broker shorthand and consumer slang described the same products in different words, and any slip into advice, rates or approvals was a compliance problem.
01
Mismatched audiences
The same channel served brokers wanting technical depth and consumers wanting reassurance in plain language.
02
Wrong product, lost lead
Generic chatbots mapped refi and no money down to the wrong loan type and lost the enquiry.
03
Strict boundaries
No specific quotes, no financial advice and no approval promises, on every single reply.
The solution
Classify, translate, retrieve, and guard every answer.
Each enquiry is classified as broker or consumer on arrival and the agent adjusts depth and tone, then routes the conversation to a direct answer, follow-up context, retrieval-grounded product matching or a graceful out-of-scope redirection.
A translation layer maps slang and shorthand onto the real product catalogue so the agent understands what both audiences are actually asking for, across more than thirteen indexed products.
A guardrail layer intercepts any request for advice, specific rates or approvals and responds safely. Responses come from a premium reasoning model while routing and retrieval run on a faster model for cost control, with a multi-model fallback chain behind both, and long conversations summarise themselves to stay coherent.
Broker and consumer classificationRetrieval-grounded product matchingColloquial-to-product translation layerCompliance guardrail layerPremium reasoning model with a fast routing modelMulti-model fallback chainAutomatic conversation summarisation
13+Loan products indexed with colloquial-to-product translation across broker and consumer vocabulary
2Audiences routed within a single agent, live in production
0Quotes, advice or approval promises exposed, enforced by the compliance guardrail layer
100KToken context kept coherent with automatic conversation summarisation and a ten-model fallback chain
Questions
About this project.
It classifies each enquiry as broker or consumer on arrival and adjusts depth and tone: quick technical answers on eligibility, documentation and edge cases for brokers, plain-English guidance for consumers.
A compliance guardrail intercepts requests for quotes, financial advice or approval promises and redirects them, so no reply exposes a rate, advice or an approval.
A colloquial-to-product translation layer maps shorthand such as refi or no money down onto the real catalogue of more than 13 loan products, backed by retrieval-grounded product matching.
