Voice AI for Outbound Lending Sales

Overview

I led the development of an outbound Voice AI sales channel for personalized lending offers at a large digital commerce platform. The agent contacts users with an available offer or prior interest, understands their needs, answers objections, and guides them toward completing the process in the official app.

The work covered the full operating chain: audience preparation, campaign rules, outbound dialing, real-time voice, structured outcomes, automatic evaluation, and Salesforce integration.

From discovery to production

I started by studying human sales calls, training material, product rules, and quality criteria with the business team. I translated those findings into product requirements and a governed conversation flow, then ran controlled pilots with real users.

Actual calls drove the iterations. We refined discovery questions, objection handling, closing, voice selection, interruption behavior, and latency before moving to higher-volume production traffic in Mexico.

Adapting outbound sales to Mexico

Mexico introduced a behavior we had not encountered in the same way in Argentina: automated call-screening assistants could answer first and start bot-to-bot conversations.

Automatic evals surfaced and measured the pattern. I added a dedicated node to the ElevenLabs workflow and adjusted the prompting so the agent could recognize the scenario, follow a controlled exit, and avoid treating it as a conversation with the account holder.

Localization covered the opening style, pacing, vocabulary, turn-taking, objection handling, and voice selection so the complete interaction fit the market.

Production architecture

Genesys manages campaign orchestration and dialing, while ElevenLabs powers the real-time voice agent. Before each call, deterministic preprocessing prepares names, phone numbers, dates, amounts, rates, and offer context so the model can focus on the conversation.

After each call, a webhook sends the event to an internal processing layer. The outcome is normalized and evaluated automatically before the corresponding result is written to Salesforce. Those structured results support recontact strategy, funnel measurement, quality review, and follow-up workflows.

Safety rules apply throughout the flow. The agent speaks only with the account holder, never requests credentials, does not invent or modify offer terms, respects opt-out requests, and keeps sensitive actions inside the official application.

My role

As lead developer and technical owner, I coordinated directly with the business team, translated operational needs into product and technical requirements, and developed the conversational agent. I also built the telephony and post-call integrations, evaluation and observability flows, and led the system from discovery through production traffic.

Results

In a full-month run, the agent covered a portfolio of more than 13,000 users and executed roughly 100,000 dialing attempts. It reached over 3,400 confirmed positive contacts, about 26% of the assigned portfolio, without requiring a representative during the main conversation. Around 16% of those contacts expressed interest.

The run demonstrated operational scale, strong post-contact efficiency, and the capacity to handle structured outbound sales conversations. Incremental originations and financial ROI remained a separate evaluation track.