Gynger Offer Agent Case Study

Overview

Offer Agent

Replacing static transactions and credit underwriting with an AI-native agent

Designing a conversational agent that coordinates buyer, vendor, and ops

I designed Gynger's first AI-native product, a conversational agent that replaces the static checkout form, guides buyers through credit application, and routes exceptions to the right human at the right time. The work included conversation architecture, hard guardrails, and the system prompt itself as a design artifact.

Discovery

Moving from a form that stalled deals to an agent that closes them

The existing checkout was a static credit application. Buyers got an offer link, navigated the form, and either completed it, abandoned, or got stuck and required customer success to chase them. CS involvement on every stalled application meant revenue was bottlenecked by headcount willing to follow up.

Three parties, one experience, information isolation by design

The buyer experiences a conversational agent guiding them through options and application, seeing only what's relevant to their decision. The vendor receives templated Slack messages for exceptions and status updates. Gynger ops handles credit decisions on complex cases via a structured context packet. Each party sees a different slice of the same conversation.

Designed conversation as a state machine with classified intents

Every message gets classified into one of five intents — qualifying, standard, exception, sensitive, or meta — and routed accordingly. Qualifying steers to the best product. Exceptions escalate to vendors via templated Slack messages, never LLM-generated. Sensitive data routes to UI form fields, never collected in chat.

Outcomes

Outcomes & learnings

The offer agent is currently in a beta release, so data is still being collected on its efficacy. The learnings of the project can be extended to similar projects when designing agentive surfaces.

In beta, early signals are strong on completion time and CS reduction

Beta targets: sub-45-minute completion from offer link to signed agreement, 70%+ of started applications reaching underwriting, zero CS-required reviews from missing agent guidance. Early signal: completion times significantly faster than the form-based flow; CS involvement in deal closure has dropped.

AI design problems are behavioral, not visual, and constraints are features

Tone mattered more than any layout decision. The agent's first response sets the trust threshold for the entire conversation. Templated vendor messages felt limiting at first; they became the feature that made ops trust the agent. And designing for AI requires comfort with uncertainty: you're designing a distribution of outputs, not a deterministic path. The work moves from specifying screens to shaping behavior.

Dylan Ginoza