AI Copilot

Infinitus's AI agent handles millions of healthcare phone calls on behalf of providers, verifying benefits, scheduling, and retrieving patient information. As the system scaled, operations teams needed a way to monitor, review, and intervene in AI-handled calls without slowing the throughput that made the product valuable.

The challenge: design a Copilot interface that surfaces the right information at the right time, builds operator trust in autonomous decisions, and enables fast, confident action when human judgment is needed.

When I joined Infinitus, the AI Copilot was a functional but sprawling internal tool that had grown organically alongside the product. No dedicated designer had shaped it. Below is what it looked like on day one.

Discovery was a structured research program across the five user groups who touch the AI Copilot: benefit verification specialists, A/R claims specialists, team leads, supervisors, and managers. I led every interview personally, working from role-specific discussion guides that opened on day-to-day workflow and laddered up to what “confident handoff to AI” meant for each role. The sessions were semi-structured: I wrote a guide tailored to each role, built new ones as needs emerged, and followed any thread worth digging into wherever it opened mid-interview.

To go beyond what people could articulate on a video call, I flew on-site to a customer’s headquarters for multi-day shadowing with their benefit verification team. Watching the work in the room (the second screen they kept open, the spreadsheets they pasted from, the colleagues they tapped on the shoulder) surfaced friction no remote interview would have caught.

What I heard back wasn’t one uniform need. Specialists wanted clarity in the moment. Team leads needed real-time signal on who was struggling without breathing down anyone’s neck. Supervisors and managers needed an oversight surface, so I designed a separate admin view for reviewing operator work, tracking performance, and turning Copilot call data into team-level insights they could act on.

Workflow mapping then revealed how the existing tooling forced operators to jump between multiple internal dashboards to piece together call context, creating lag in escalation decisions and error risk in high-stakes healthcare work. From there, I led concept design and rapid prototyping cycles with engineering and ML partners, pressure-testing assumptions about when AI confidence scores were meaningful vs. misleading to non-technical operators.

These sessions produced more than notes. I synthesized the learnings into end-to-end user flows: a shared artifact mapping how each role moves through the Copilot across its rollout phases, from new-user setup through the pilot and post-pilot updates. A flow like the one below gave customers and internal partners something concrete to align around, turning scattered interview insights into one picture of the workflow, its pain points, and how the Copilot would meet each need.

Qualitative

• Only 1.7% of failed calls were attributable to the AI itself. The rest were user-abandoned, bad data, or payer system issues, evidence that the Copilot’s confidence signals were honest about when to escalate and when to keep going.
• A small ops team worked ~13K payer tasks daily. 18 active users averaged 8.4 successful calls each per day, with the Copilot’s queue prioritizing what needed human attention instead of forcing review of every call.
• Scaling didn’t require scaling headcount. Throughput and revenue grew on the same operator footprint, the original design bet behind the Copilot.

AI Copilot: Live Assist. The current Copilot enables oversight of calls after or during a review pass. The next evolution removes that lag entirely, collapsing the gap between AI action and human judgment:

• Real-time call monitoring & live transcription. Operators see and hear what the AI agent is doing as it happens, not after the fact.
• AI recommendations in an ambient panel. Suggested next-best actions, flagged moments, and contextual notes surface inline so operators can act without leaving the call view.
• Listen in, or jump in. One tap to silently audit a live call; one more to take over when human judgment is needed.
• Connected to Infinitus’s intelligence layers. Insights from the broader platform (payer behavior, historical outcomes, claim patterns) surface where the work is happening, not in a separate tool.
• Validated with the customer advisory board. Designed and prototyped in close collaboration with SVPs and VPs of Patient Access Operations, and validated at the executive level as the direction they want to evolve toward.