Zero-Tolerance Policy Enforcement

A bad call here doesn't cost a sprint

In a regulated call center, a missed compliance violation isn't a UX papercut. It's a fine, a lost license, or a headline. Leadership was flying blind: quality teams manually reviewed roughly two percent of calls, violations surfaced days later, and by the time anyone acted the damage was done. The real problem wasn't effort. It was a system blind to risk at scale.

Why manual auditing kept failing

• Sampling missed the long tail. The two percent reviewed was never the two percent that mattered.
• Feedback came too late to change behavior. Agents learned about a violation days after the call.
• Auditors drowned in low-risk calls and had no way to triage toward genuine risk.
• Leaders managed reactively, explaining incidents after the fact instead of preventing them.

My role as the design lead

I led design end to end: the research with auditors and agents, the information architecture for surfacing risk, the core interaction model, and the working relationship with the ML team that decided what the model could and couldn't be trusted to flag. My job wasn't to decorate a dashboard. It was to design the judgment layer between an imperfect model and a human who had to act on it.

• Ran contextual interviews with auditors to map how they actually triage risk.
• Defined a risk-scoring surface that ranked calls by likelihood and severity.
• Designed the human-in-the-loop review flow so people could confirm, dismiss, or escalate fast.
• Set the trust model: where the AI decides, where it suggests, and where a human must sign off.

Design as risk management

The core insight: not all violations are equal, and not all model confidence is equal. So the interface had to communicate two things at once — how risky a call was, and how sure the system was. We designed a triage queue that pushed high-severity, high-confidence calls to the top, with clear states for the gray-area cases that genuinely needed a human.

We deliberately avoided full automation. A model that auto-closes calls feels efficient until it's confidently wrong about a serious violation. The design kept a human accountable for every high-stakes decision while letting the AI remove the grunt work of finding the needle.

From reactive to preventive

Coverage went from a sliver to every call. Violations that used to surface days later now flag in near real time, so a team lead can coach an agent the same day. Within two quarters, reported compliance incidents dropped sharply, and leadership shifted from explaining incidents to preventing them.

• 100% call coverage replaced ~2% sampling.
• ~73% reduction in compliance incidents over two quarters.
• Audit team time redirected from finding violations to resolving them.

The leadership lesson

The temptation with AI products is to maximize automation and call it progress. The leadership call here was the opposite: to deliberately keep humans in the loop where the cost of a confident error was unacceptable. Designing that boundary — and getting product, engineering, and compliance to agree on it — was the real deliverable. The interface was just where the agreement became visible.

Executive Summary

Success Metrics

Stakeholder Map

Research Plan

Research Insights

Jobs to Be Done

Opportunity Map

Design Principles

Workflow Architecture

User Flow

Information Architecture

Concept Exploration

Wireframes

Design System

Final Design

Usability Testing

What I'd Do Next

Future-State AI Vision