Enterprise Risk Intelligence · AI Governance

A compliance system that finally sees risk at scale.

Fusion BPO audited under 15% of calls by hand in a high-liability environment. Violations surfaced days late. AI had been tried and rejected — QA teams wouldn't trust black-box flags. In four weeks I led the UX and governance design that turned sampling into continuous, explainable monitoring of 100% of calls. The breakthrough wasn't accuracy. It was explainability.

See the impact View the work
100%
Calls monitored (from <15%)
90%
Adoption among non-tech QA
+30%
Compliance lift
$2.25M
Projected annual ROI
Scroll
ztp-intel.app/governance Risk overview Audit sheetFlagged callsEscalationsTraining triggersReports Risk overview Live · 1,284 calls today Coverage100% Open · high sev.9 Accuracy95% Feedback4 min Violations by severity · 14 days Flagged queue HIGH #A-4821 · abusive languageconf. 0.94 · 2m ago MED #A-4820 · missed disclosureconf. 0.71 · 6m ago LOW #A-4817 · tone / sentimentconf. 0.58 · 11m ago Review queue →

The governance view — 100% coverage, violations triaged by severity and confidence, feedback in minutes.

01 — The Trap

Zero-tolerance policies that existed on paper, and nowhere else.

The organization wasn't operating inefficiently. It was operating blindly.

In a high-volume, high-liability contact center, policies against abusive language, regulatory breaches and brand conduct were strict — on paper. Enforcement ran on manual audits, random sampling, delayed feedback and subjective judgment. That combination created blind spots where violations went undetected and leadership could only react after the damage was already done.

The exposure stayed invisible until it became a legal or revenue problem: roughly $500K a year in legal liability and up to $1M at risk from churn and brand erosion. And even a technically strong AI would fail here — non-technical QA teams feared false positives and personal accountability, so trust was the real gate, not accuracy.
02 — The Insight

Not an efficiency problem. A risk-visibility and decision-confidence problem.

That reframe moved the project from "automate audits" to "build governance infrastructure." Discovery ran tight and deep under a four-week clock: 12 QA analysts shadowed, 18 past violation cases walked through in retrospect, and dedicated AI-trust and mental-model sessions. Three insights set the entire direction.

Ambiguous flags killed action instantly

An AI flag with no reasoning triggered immediate distrust. Analysts would either ignore it or rubber-stamp it — both useless. "Why" wasn't a feature request; it was the precondition for any action at all. Reason, evidence, severity and confidence had to ship together or the system was dead on arrival.

Adoption hinged on workflow fit, not feature depth

The 50–60% resistance was rooted in false-positive fear and accountability anxiety, not missing capabilities. A tool that disrupted the existing QA flow would lose, however clever. Seamless integration beat breadth every time.

Managers needed patterns, not just incidents

To shift from lagging to proactive governance, QA managers and VPs needed severity-based trends and live escalation visibility — the difference between counting fires and preventing them.

Synthesis: trust and explainability were the highest-leverage lever. I bet on them over feature breadth — the only path to 90% adoption in a non-technical environment.
The systemic shift BEFORE · BLIND Coverage<15% Review time20 min/call Feedback delay7+ days Accuracy75% Adoption50–60% resist AFTER · GOVERNED Coverage100% real-time Review time−40% effort Feedback delayMinutes Accuracy95% Adoption90% trust
FIG 01 The before/after that mattered to leadership — every line a blind spot closed.
Two personas, two layers: execution and governance QA QA Analyst PRIMARY · non-technical, time-pressed "If I can't see why it flagged, I won't act on it." GOALFast, fair violation decisions PAINAmbiguous flags, accountability fear NEEDContext, confidence, speed → Drove: progressive disclosure + the 4-part signal VP QA Manager / VP SECONDARY · governance oversight "I need patterns and escalation, not a pile of incidents." GOALReduce risk exposure, improve trends PAINLagging indicators, manual escalation NEEDSeverity, patterns, governance visibility → Drove: severity dashboards & live escalation
FIG 02 Personas as decision drivers — the analyst's distrust shaped the signal; the manager's need for patterns shaped the dashboards.
STAKEHOLDER INTERVIEWS·CONTEXTUAL INQUIRY·12 ANALYSTS SHADOWED·18 VIOLATION RETROS·AI-TRUST TESTING·RICE PRIORITIZATION· STAKEHOLDER INTERVIEWS·CONTEXTUAL INQUIRY·12 ANALYSTS SHADOWED·18 VIOLATION RETROS·AI-TRUST TESTING·RICE PRIORITIZATION·
03 — The Build

An AI risk-intelligence layer that explains itself.

The architecture rested on one principle that came straight out of research: never show a verdict without its reasoning. Every flag carried four things together — reason, evidence, severity and a confidence score. That single structure is what turned AI from a surveillance tool the team feared into governance infrastructure they trusted.

1

Live risk detection

Continuous AI analysis of every call at 95% accuracy, eliminating the sampling blind spots that hid risk until it became a lawsuit.

2

Explainable signals

Each violation arrives with its reason, the supporting evidence, a severity level and a confidence indicator — the "why" that makes a flag actionable.

3

Adaptive governance

Auto-triggered training, policy updates and escalation rules close the loop, turning each detection into a continuous-improvement signal for ops.

Card sort → the platform IA QA analysts grouped the tasks they actually do. Their mental model became the navigation. TASKS SORTED BY ANALYSTS review flagged call override false + see risk trends escalate to manager create audit sheet update policy rules export report assign training Sorted into create / view / manage groups RESULTING NAVIGATION Dashboard Audit Sheet Flagged Calls Audit Trails Escalations Reports Users & Roles Settings PER-ITEM ACTIONS (FROM SORT) CREATE VIEW MANUAL AI Clarity over complexity: non-technical users navigate by the task they're doing, not by system architecture.
FIG 03 The card sort that set the IA. Analysts' task groupings became the navigation, keeping a complex AI tool legible to non-technical users.
The signal that earned trust: reason + evidence + severity + confidence Progressive disclosure — summary first, deep-dive on demand. CALL #A-4821 · 06:12 · AGENT 23 Customer · 04:31"This is the third time I've called." Agent · 04:51 — EVIDENCE "Honestly, that's not my problem to fix." flagged moment · brand-conduct breach ▶ Play 04:45 REASON Dismissive refusal to assist with a customer issue. Maps to policy ZTP-04 · Brand conduct & empathy Progressive disclosure keeps the table fast; detail is one tap away. SEVERITY HIGH Zero-tolerance tier CONFIDENCE 0.94 high — safe to act YOUR DECISION Confirm & coach Override (false +) AUTO-GOVERNANCE Agent notified in < 5 min Coaching session auto-queued Override retrains the model
FIG 04 The analyst's core screen. The four-part signal on the right is the entire trust thesis, made literal.
Why explainability won the roadmap (RICE) Reach × Impact × Confidence ÷ Effort — the highest-leverage bet under a 4-week clock. Explainable signals 96 Severity-based dashboards 72 Auto-training triggers 58 Multi-language support 24 · cut
FIG 05 Ruthless scoping. Explainability scored highest; multi-language was deliberately cut to ship in four weeks.
From audit table to trusted signal: three fidelities Each round tested with QA analysts; trust was the metric, not task completion. LOW-FI · just a table Problem: flags with no "why" → distrust Analysts wouldn't act on bare flags. MID-FI · + reason & severity HIGH reason: dismissive refusal… Better — but is the AI sure? Missing: confidence + evidence. HIGH-FI · the 4-part signal HIGH conf. 0.94 reason + ZTP-04 mapping ▶ evidence audio · 04:45 Confirm Result: 90% adoption — analysts trust it.
FIG 06 The fidelity progression. Each round added a piece of the trust signal until analysts would act — the path from bare flag to 90% adoption.
Design system: severity-driven hierarchy Principles — clarity over complexity · severity-driven hierarchy · progressive disclosure. SEVERITY TOKENS HIGH MED LOW Color = urgency. Attention goes where risk is. CONFIDENCE INDICATOR 0.94 The AI's certainty, shown — honest about doubt. ACTION STATES Confirm & coach Override Every flag has a clear, two-path decision. AUDIT TABLE Dense data, scannable — progressive disclosure built in.
FIG 07 The component library. Severity tokens, confidence indicators, action states and the audit table — each encodes the trust-first principles.
04 — The Proof

The honest version of the numbers.

Straight talk: the operational and adoption gains were measured in the beta. The financial figures are projections. And usability landed at 69.25 — a v1 baseline built under extreme time pressure, with a clear target of 80+ in the next iteration. I'm showing it as a roadmap, not a trophy, because that's what it honestly is.

Measured in the beta

+40%
QA efficiency gain — review time cut as analysts shifted from listening to acting on signals.
90%
Trust & adoption among non-technical QA users — the metric the whole bet rode on.
+30%
Compliance lift, with coverage up from <15% to 100% of calls.

Projections & targets

$2.25M
Projected annual ROI — avoided penalties + churn reduction
+15%
CSAT improvement (projected)
69 → 80+
SUS baseline today, target next iteration
Usability: a deliberate v1 baseline, with a known next move 69.25 in four weeks under compliance pressure. The gap is integration (Q4/Q5) — and it's the iteration plan. 69.25 NOW · v1 80+ TARGET · v2 0 100 · Excellent "Good" begins at 71 — one focused iteration away.
FIG 08 SUS framed honestly as a trajectory: a strong v1 under constraint, with feature integration as the named lever to clear "Good."
The number I'd stake the project on is 90% adoption in a population that started at 50–60% resistance. In an enterprise AI rollout, getting non-technical users to trust the system is the whole game — and that came directly from the explainability bet.
05 — The Call I Had to Make

Spend the budget on explainability, not raw accuracy.

"

A more accurate black box still gets ignored. I forced explainability as non-negotiable — even at the cost of headline accuracy gains.

The tempting path was to chase a higher accuracy number, because it's easy to put on a slide. But early trust assessments were unambiguous: non-technical QA users would not act on flags they couldn't understand, no matter how accurate. So I made explainability the non-negotiable core — reason, evidence, severity and confidence on every signal — and defended it to the C-suite in their own language: every undetected violation carried real legal and churn exposure, so adoption was the ROI, not a soft extra.

Two more calls made the four weeks possible. I pushed for a staged beta over a big-bang launch, controlling false positives first and earning leadership confidence with Week 2 metrics. And I scoped ruthlessly — cutting multi-language and edge cases to ship a trustworthy single-language core on time. Those trade-offs are exactly why the SUS sits at a v1 baseline; I traded polish for a system people would actually adopt under deadline.

What I'd do differently

Tune per-persona earlier to push SUS past 80 — the Q4/Q5 integration scores were mixed and foreseeable. And run legal co-design workshops up front, so regulatory constraints shaped the design from week one instead of being reconciled later.

The Role

Client — Fusion BPO Services (large-scale contact center operations)
Role — UX Leader: strategy, research, experience architecture, trust & explainability governance
Timeline — 4 weeks, strategic discovery to live beta
Team — Me (UX ownership), 2 Sr. Product Designers (UI execution), 1 UX Researcher (ops), with AI Engineering, Ops & Legal collaborators
What I owned — End-to-end research strategy on AI trust and adoption, system-level experience architecture across analysts, managers and agents, the trust/explainability metrics, translation of legal and regulatory risk into prioritized UX requirements, and the executive narrative shift from efficiency tool to enterprise risk-control platform.