An AI copilot healthcare agents actually trust.

Healthcare agents were drowning in software instead of listening to patients.

I led the end-to-end design of a workspace for agents who schedule appointments, manage prescriptions, and handle billing over the phone. First I consolidated four fragmented systems into one canvas. When testing showed that wasn't enough, I evolved it into an AI copilot that listens to the call and does the searching so the agent can focus on the patient.

It shipped as an MVP pilot, so I have no production metrics — and won't invent any. What follows is the research, the screens, and the pivot, built around one non-negotiable: the AI recommends, the human decides.

Nextiva brought its unified CX platform into healthcare, where a "customer" is a patient and a mistake has consequences most SaaS never sees. I owned interaction design and strategy for appointment scheduling, rescheduling, prescriptions, and bill pay from research synthesis through the high-fidelity AI model.

Every minute an agent spends hunting through screens is a longer hold, a longer call, and a higher chance the patient calls back. Healthcare was Nextiva's wedge: a high-stakes vertical where a unified, careful AI workspace could beat legacy telephony bolted onto legacy EHRs under HIPAA, integrations we didn't control, and agents measured on speed who can't afford to be wrong.

Industry figures cited from Nextiva's published healthcare research — market context, not results from this project.

Lead Product Designer the single design owner end-to-end, alongside Product, Engineering, and a dedicated UX Research team.

A dedicated UX Research team led a mixed-methods study; I sat in on sessions, pressure-tested findings, and turned insight into decisions.

Agents wanted to help people, and the software kept getting in the way.

• Constant system-switching across EHRs and internal tools to finish one request.
• Fragmented patient info the full picture never lived in one place.
• The hold button as a crutch agents put patients on hold just to go find things.
• Call-backs as failure many requests couldn't close in one call.
• They wanted one workspace, not another tab to alt-tab into.

• Bring fragmented patient context into one place at the moment it's needed?
• Make scheduling feel like a decision, not data entry?
• Use AI to remove the searching without removing the agent's control?

I mapped what actually happened on a call every branch, every screen an agent had to leave to go look something up. A single "book an appointment" request could touch identity, insurance, availability, and location, each in a different place.

flow mapping — scheduling journey, per intake path

Everything in one place, with a structured flow: specialization, provider, consultation type, location, date, slot. Rescheduling reused the pattern with the current appointment pinned.

It removed the alt-tabbing. It also revealed the ceiling of "organize the chaos better."

Tidy forms are still forms. The agent was still the search engine reading availability, typing while the patient waited. Here's the actual research readout that ended this iteration:

When every participant hits the same structural wall, that's not a preference it's a signal. It told me the next move wasn't a better form. It was a different kind of help.

Don't make the search faster do the search for them.

The product evolved from a manual workspace into an AI-assisted one. NextIQ listens to the live transcript, identifies intent in real time, and surfaces the appointment, prescription, or billing action the agent would otherwise go dig for.

The copilot never took the wheel. The agent always reviewed, confirmed, or modified every recommendation. That was the design, and the thing I fought to keep see Section 17 for the disagreement it caused.

I moved low fidelity to high, deliberately flows and component logic before pixels, in a tight loop with Engineering on what the AI could reliably do.

nextiq for appointments component & state ideation

That process led to four real screens from the same call — the copilot's lifecycle, shown live rather than described.

live mode call transferred from the virtual assistant

HIPAA verify · intent detected · 96% confidence

suggested action · pre-filled scheduling widget

appointment confirmed · 4 automations running

nextiq · ai copilot demo

Demo video goes here

Drop a screen recording in as assets/copilot-demo.mp4 and it plays here automatically.

Eight principles kept a powerful feature from becoming a reckless one.

The hardest: explainability under time pressure. Confidence became a glanceable signal; the "why" became a few chips — see Section 14's 96% confidence panel.

Naming what I gave up, and why the trade was worth it, is how I keep a design honest.

We tested both acts against realistic and messy scenarios — a patient who changes their mind, no available slots, a request a front-desk agent shouldn't complete at all. The manual version taught us its ceiling; the copilot version tested as the part of the job agents liked least, gone.

research readout — other feedback

I watched for the failure I feared most — agents rubber-stamping suggestions. Designing the confirm as a deliberate, evidence-backed act was a direct response.

A pilot with a handful of agents doesn't prove the design holds at enterprise scale. What I addressed, and what's still open:

• EHR variance — the graceful-failure principle exists specifically for this: when the AI can't reliably read a given EHR, it says so and steps back to manual.
• Language and accessibility — the pilot ran English-only; multilingual handling and screen-reader support for confidence chips are open V2 questions.
• Volume — at scale the risk isn't the UI, it's intent-detection latency and accuracy under load, which needs tighter ML/Engineering monitoring.

Testing also forced states I hadn't drawn yet — cancellations, no-shows, re-verification — before they became production bugs.

The product shipped as an MVP pilot — a scoped, real-world test, not a full rollout. Production metrics don't exist yet, and I won't manufacture them. "Launch" meant readiness: a coherent, safe, buildable experience that clinical stakeholders were willing to put in front of real agents.

Honest, qualitative, and enough to justify the next investment:

• Consolidation is table stakes, not the win. The real leverage was changing who does the searching.
• In AI products, trust is the feature. The moment agents could override the copilot, they started relying on it.
• Design the AI's humility. How it admits doubt mattered more than how clever it was on a good day.
• Test the model, not just the screens. The manual scheduler looked fine and tested poorly, because the problem was structural.

My job wasn't to design the smartest AI in the room — it was to design the relationship between a stressed human and a fallible machine, where being wrong isn't an option. I made the case for a pivot when a tidy first version tested well enough to ship and not well enough to matter, and held the line on human control when a stakeholder pushed for more automation than I believed was safe.

A V2 starts with the real practitioner calendar agents asked for, and billing and insurance made permanently visible. The principle carries into everything I design now: for enterprise AI, copilot, never autopilot.