Jordan Campbell · Director of Product Design & AI Operations at Inbox Health

Selected work

Inbox Health AI Support

Designing AI chat and phone customization, an in-app support assistant, and Voice AI configuration.

Senior UX Designer → Director of Product Design → Director of Product Design & AI Operations · 2024-Current · Inbox Health

Since June 2024 I’ve designed AI and conversational features at Inbox Health, aimed at self-service support. I designed, 0 to 1, the customization for our AI support chat and AI phone calls, an in-app AI support assistant for billers and practice staff, and Voice AI configuration. Support staff review a response before it is sent.

Results

The difficult part

What shaped the work

Patient-facing healthcare sets the rules: HIPAA access controls, audit logs, and a person reviewing AI output before it reaches a patient. Guardrails sit in a master prompt that a practice can override or extend, and call analysis sends only the customer side of a call to the model.

Who I worked with

I mapped how the support teams work before designing for them, from time-study data, ticket volumes, SOPs and Intercom root causes. Support staff review AI responses before they are sent, and internal team members use the assistant themselves.

Customization

Practices set the rules, a person reviews

I designed, 0 to 1, the customization for our AI support chat and AI phone calls. Practices set scripts, rules, escalation to a human, and hours. A side-by-side live chat preview tests scenarios, including against a real patient record. Support staff review a response before it is sent, and guardrails sit in a master prompt that can be overridden or extended.

Voice AI

Phone calls a practice can set up itself

For Voice AI I designed custom greetings before and after caller verification, a per-number AI answering toggle, and self-serve phone-number provisioning.

The Assistant

Help inside the app, with a way to a person

I designed an in-app AI support assistant for billers and practice staff, with starter questions, settings awareness, and escalation to a human agent. Alongside it: AI ticket triage that filters out bot-resolved tickets, an Intercom Fin help-center chat scoped to one client’s knowledge base, and natural-language billing-activity queries for operations leads. Internal team members use it too.

Where It Started

AI features since June 2024

The first AI features I designed were LLM-suggested replies for patient-support agents, AI call transcription with real-time response suggestions, and per-client AI prompt configuration with simulation. I also designed an in-app AI response tool that replaced static documentation, so users learn by asking instead of reading.

Evaluating the AI

Blind replays and a parity scoreboard

I evaluate AI automations with blind replays. Past calls run blind through the billing-cycle ticket-audit flows, each pass is scored on a parity scoreboard, and the fixes feed into the next pass, including narrowing rules that over-corrected on the pass before.

After the Gong sunset I rebuilt the company’s call trackers on an LLM: an n8n workflow checks each recorded call against a written rubric with Gemini or OpenAI models and posts matches to Slack. Only the customer side of each call is sent to the model. On a real week of calls it matched Gong’s precision at 100%. It is still in development.

Demos

For client and sales presentations

I designed conversational AI demos for client and sales presentations: the Inbox Assistant demo, with thinking and error states and a reasoning-effort slider; a Voice AI demo in the patient experience prototype; and an AI billing chat with Stripe payments inside an EHR partner’s patient app, packaged for that partner’s demo at a national health IT conference.

Research Behind the Roadmap

How the support teams actually work

I ran the support-automation research behind the company’s AI agent roadmap. I mapped how the CSS, PBS, and tech-support teams work from time-study data, ticket volumes, SOPs, Intercom root causes, and product analytics, and identified 37 automation candidates, each tied to a 2026 OKR. One of those OKRs is a 30% support-deflection target, which is a goal, not a result.

Next case study: Inbox Health Partner API Docs