Case study · Design Challenge · Speculative design · 2026

NovaCorp · turning messy invoice data into recovered cash, with AI as a teammate.

Enterprise platform that helps procurement and finance teams find financial leakage hidden in supplier invoices. From raw data ingestion through compliance scoring, exception handling, and executive reporting — designed so that an AI agent does the heavy lifting transparently.

RoleUX / UI Designer
Scope4 hi-fi screens + 4 states
DomainProcurement & Finance · B2B SaaS
Timeline5 days

Large enterprises lose millions to invoice errors they never even notice.

NovaCorp processes thousands of supplier invoices every month across multiple countries and currencies. Each one needs to be validated against signed contracts, approved rate cards, and commercial agreements. Today the process is largely manual — analysts compare line items by hand, exceptions slip through, and the company quietly pays for surcharges that contracts don't authorize.

The platform's job is to find the money the company is bleeding — and to do it without anyone needing to become a data engineer.
$1.28M
Total financial exposure across 247 invoices in a typical quarter
3
Distinct user types — procurement, finance, operations — each with different jobs-to-be-done
8
Categories of error to surface, from currency mismatch to outdated rate cards

Three users, three different relationships with the same data.

A single dashboard can't serve all of them equally well. The design picks the right primary view per persona and lets the other two drill in.

Persona 1
Maya · Procurement Manager
"Help me know which vendors I need to push back on this week — and how much money is at stake."
Needs Triage vendors by exposure · Build credible recovery cases · Spot systemic vs one-off patterns
Persona 2
David · Finance Controller
"Give me numbers I can defend in front of the audit committee, and tell me where the trend is going."
Needs At-a-glance financial KPIs · Defensible breakdown · Trend over time
Persona 3
Sarah · Operations Analyst
"Let me trust the data before anyone else uses it — and let me drill into the invoice when something looks off."
Needs Data quality verification · Audit the AI pipeline · Drill from exception to root cause

Three principles that resolved every disagreement.

A data platform for time-pressured enterprise users has to feel different from a consumer app. These three guardrails kept the design honest.

01

Clarity over decoration

No illustrations, no flourish. White space exists only to separate information. Every pixel earns its place by serving a decision the user has to make.

02

Progressive disclosure

Summary at the top, drill-down within one click. Every screen answers a single question and points the eye at the next action.

03

AI is a first-class teammate

The agent is visible, transparent, and never acts silently. It shows evidence and confidence before offering any action. Purple is its consistent signature so users can always tell who did what.

Compliance Overview — the morning triage view.

Maya opens this on Monday with 30 minutes before her first call. The screen has to answer "where do I focus this week?" instantly.

Phase 2 · Audit

5 KPI cards · AI insight panel · waterfall · vendor ranking · sortable table

Above the fold: compliance percentages, total exposure, net leakage. The AI surfaces a $48K recoverable pattern before the user even scrolls. The vendor table is sorted by exposure descending so the worst offender lands at the top of the eye.

Five filter dimensions

Vendor · Date · Geography · Contract · Service Level — the brief's required filter set, all visible above the table.

Waterfall chart

Tells the financial story in one image: $430K overcharges − $228K undercharges − $18K AI-resolved = $184K net leakage.

AI insight panel

Pre-surfaced pattern: "17 invoices show the same surcharge tier error · $48,210 recoverable · 84% confidence." Two action paths.

Screen 1 — Compliance Overview

Click to open the full 1280px hi-fi mockup in a new tab.

Open Screen 1 →

Data Quality Report — the agent shows its work.

Sarah lives here. She'll trust the data only after she's seen exactly what the cleaning pipeline did. So the AI doesn't just run — it shows itself running.

Phase 1 · Cleaning

Before / after score · 6-step pipeline · issue breakdown · ingestion log

The pair of circular gauges (67 → 94) is the headline. The pipeline below it is the receipt. When the agent stops at step 4 for vendor harmonization, the amber color and "Needs review" badge make it impossible to miss what's being asked of the human.

Score gauges

Two circles, one delta. The orange-to-green color shift communicates the win without anyone reading a number.

Agent paused for review

Step 4 stops on amber with two CTAs. The agent never proceeds on ambiguous data.

Five issue categories

Missing values, duplicates, schema mismatch, currency, invalid references — each shown before vs after.

Screen 2 — Data Quality Report

Includes a state variant showing the agent running live.

Open Screen 2 →

Invoice Detail — the math is done for you.

When a vendor row is clicked, the user lands here. The job is to make a recovery decision: dispute, approve, escalate. The page is built to let you do that in 30 seconds.

Phase 2 · Audit

Invoice header band · Invoiced vs Expected · exception cards · AI batch recovery

Every line is shown side-by-side: invoiced vs expected vs deviation. $0 deviations recede in green; positive ones jump out in red. The AI panel connects this single invoice to 17 siblings with the same error — turning a one-line dispute into a $48K systemic recovery.

Exception card structure

Four fields per exception: deviation type · financial impact · root cause · recommended action.

17-invoices-like-this

The AI extends a single dispute into a batch — the most differentiated moment in the design.

Activity log

Append-only trail of every action — human and AI — for audit defensibility.

Screen 3 — Invoice Detail

Single source of truth for any disputed invoice.

Open Screen 3 →

Root Cause Analysis — where do we focus to fix the most?

The strategic view. The Pareto chart shows that 4 of 8 exception categories drive 81% of all leakage. The user can stop solving symptoms and start solving causes.

Phase 2 · Audit · Strategic

Pareto chart · 8 category cards · 3-tier AI insights · 4-quarter trend

The chart's cumulative line crosses 81% at category 4 — the textbook 80/20 break, explicitly labeled. Below it, three AI insight cards translate the data into action: a systemic issue, a predictive alert, and a quick win. Each is colored by what it asks of the user.

Pareto with cumulative

Two encodings, one chart. The bars rank by impact; the line shows when cumulative coverage hits 80%.

Three-tier AI insights

Systemic (red) for rollbacks · Predictive (amber) for what's coming next quarter · Quick-win (green) for instant auto-resolves.

Toggle: frequency vs $

Same data, two questions. Procurement strategy switches between counting and weighting.

Screen 4 — Root Cause Analysis

Strategic view for procurement leadership.

Open Screen 4 →

The agentic flow — three modes, never silent.

The brief asked for AI agents that "proactively identify, flag, and resolve problems." The hard part was making them trustworthy. The answer was a deliberate three-tier model so the agent's actions match its confidence.

!

Flag for human

High-impact deviations. Confidence ≥ 75%, financial impact above threshold. Always presents evidence first — never auto-acts.

Screen 4 · Systemic issue: "Rate-card v2.3 ingestion drove 73% of surcharge errors. Roll back recommended for 3 contracts."

Suggest action

Medium confidence (50–75%). Shows reasoning. Action button requires explicit human approval.

Screen 1 · Pattern: "17 invoices show the same surcharge tier error. Estimated recoverable: $48,210."

Auto-resolve

High confidence + low impact. Logged with rationale. User can audit at any time but action proceeds without blocking.

Screen 4 · Quick win: "9 outdated rate-card matches can be auto-resolved · $42K recoverable."
If it's purple, the agent did it.

Every AI-generated insight, button, and badge uses the same purple (#6D28D9). That single rule means a user can always tell — at a glance, on any screen — whether the platform did something or a human did.

A handful of consequential decisions.

The decisions that shaped the design, and why each one stuck.

Information architecture

Sidebar navigation organized by three workflow phases, not by feature.

The sidebar groups screens into Phase 1 (Cleaning), Phase 2 (Audit), and Phase 3 (Benchmarking). This matches how the platform's pipeline actually runs — and reinforces the guiding principle that benchmarking starts with data cleaning. Users always know where they are in the journey.

Visual language

ZoomInfo-inspired dark sidebar + light content area.

The brief named ZoomInfo as the design reference. The dark navigation signals "data-dense B2B SaaS"; the light content area keeps the dense tables and numeric grids legible. IBM Plex Sans (body) + IBM Plex Mono (numbers) — credible enterprise feel plus fixed-width financial scannability.

Charts

Each chart type answers exactly one question.

Waterfall for "where did the net come from?" Pareto for "which 20% causes 80%?" Score gauge for "did we improve?" The wrong chart in the wrong place hides the answer. The right chart shows it without explanation.

Agentic flow

The agent shows its evidence before offering any action.

Every AI panel leads with the data — "17 invoices · 3 contracts · 1 rate card" — followed by the financial impact, followed by a confidence percentage, followed only then by a CTA. Trust is earned through visibility, not assumed.

States

The agent refuses to operate on incomplete data.

In the error state, when 2 of 47 data sources are unavailable, the AI panel switches to a "Paused" badge with an explicit message. False patterns from partial data would be worse than no insight at all — so the design makes refusal the default.

Six things this case study is meant to show.

The brief asked for screens. The submission delivered a system — and the way that system was put together is itself part of the answer.

🎯

Information design for dense data

Five-card KPI rows, dual-encoded compliance bars, status badges with dots, sortable headers with active states, sparkline trends per row — every cell does work without crowding.

🧭

Workflow-driven IA

Sidebar phases (P1 → P2 → P3) literally mirror the data pipeline. Users build a mental model in five minutes that holds up over a year of use.

Designing trustable AI

Three-tier action model, confidence scoring, evidence-first surfacing, and a paused-during-uncertainty default. AI that knows when not to act.

🎨

Token-first design system

19 sections covering color, typography, spacing, every component, every pattern. Color contrast verified to WCAG 2.1 AA. Built so a dev team could ship without asking a single follow-up.

🪄

Beyond static screens

Four state variants (empty, loading, error, agent-running) plus a fully interactive HTML prototype with functional sorting, filtering, and a 7-step autoplay story.

📚

Documented decisions

Walkthrough doc, agentic flow one-pager, user stories, sprint plan, and a single-source design spec. The work is reproducible, not just final.

Want to dig in?

The full Figma file, an interactive prototype with click-through navigation and an autoplay story, and the design system PDF — all open.