Case Study 05 · Fintech Operations

Payment
Reconciliation
Engine

Finance operations teams spend 60% of their reconciliation time on exceptions that represent only 3% of transaction volume. This is the design process behind an AI-assisted reconciliation tool that flips that ratio.

Financial Operations Exception Management AI-Assisted Matching Data Reconciliation Ops UX

ROLELead Designer

USERSFinance Ops, Treasury

CORE PROBLEM60% of time on 3% of transactions

SCOPEProduct + AI Interaction Design

TIMELINE6 weeks

Phase 01 — Discovery

The 60/3 Problem

Through a week of process observation and workflow shadowing with a treasury operations team, I discovered the core inefficiency: time spent is completely inverse to financial impact.

Process Audit — One Week of Workflow Shadowing

I logged every task a reconciliation analyst performed across 5 days. The findings were stark: the team's workflow was organized around the wrong priority.

TIME ALLOCATION AUDIT

MATCHED transactions (97% of volume): 8 min/day to verify — automated
PARTIAL matches (2% of volume): 90 min/day — fuzzy matching, manual review
UNMATCHED transactions (1% of volume): 3.5 hrs/day — manual investigation

TOTAL: 5 hrs 8 min daily · 97% of time on 3% of transactions

ROOT CAUSE The tool shows all transactions equally. No triage. No prioritization.
DESIGN OPPORTUNITY AI handles matched + partial; human focuses only on true exceptions

Exception Taxonomy — What Actually Causes Mismatches

I catalogued every exception type from 6 months of historical data. 78% of exceptions fell into 4 solvable categories — meaning AI could resolve them without human involvement.

EXCEPTION TAXONOMY

TYPE	% OF EXCEPTIONS	ROOT CAUSE	AI SOLVABLE?
Timing difference	34%	Bank vs. internal ledger settlement date lag	Yes — auto-match with date tolerance
Amount variance <$1	22%	FX rounding, bank fee absorption	Yes — auto-approve below threshold
Reference mismatch	14%	Different payment reference IDs across systems	Yes — fuzzy match on amount+date+counterparty
Duplicate entry	8%	Double posting in source system	Yes — flag and suppress duplicate
True discrepancy	22%	Genuine error requiring investigation	No — human required

Phase 02 — Design Architecture

Exception-first interface design

The Triage Principle — Sort by Urgency, Not Chronology

Every reconciliation tool I reviewed sorted by date. I sorted by exception severity instead — unmatched and large discrepancies always at the top. This single change reduces time-to-first-action by an estimated 40%.

SORT ORDER DESIGN DECISION

EXISTING TOOLS: Sort by [date desc] — analyst must scan entire list to find exceptions
OUR APPROACH: Sort by [exception severity desc] → [amount desc] → [date desc]

PRIORITY 1: Unmatched, high value (>$10K) → immediate human review
PRIORITY 2: Unmatched, low value → AI suggestion surfaced, one-click accept
PRIORITY 3: Partial match → AI proposes resolution, human confirms
PRIORITY 4: Matched → auto-confirmed, visible but not requiring attention

RESULT Analyst sees the most important item first, every single time

Split-Pane Architecture — Table + Detail Panel

I tested three layout patterns: full-page table (no detail), modal detail (interrupts flow), and split-pane (detail alongside table). Split-pane won decisively — analysts could review details without losing context of the full list.

LAYOUT TESTING RESULTS

LAYOUT	TASK TIME	ERROR RATE	USER PREFERENCE
Full-page table	4.2 min/exception	12%	2/8 preferred
Modal detail	3.8 min/exception	8%	2/8 preferred
Split-pane (final)	2.1 min/exception	3%	4/8 preferred

Phase 03 — Design System

Warm, documentary, trustworthy

Why Warm Neutrals — The Psychology of Financial Trust

Reconciliation is about trust — trusting the numbers, trusting the match, trusting the system. I chose warm parchment neutrals with deep ink tones instead of clinical white+grey. Playfair Display for headers communicates considered, authoritative weight. The moss green accent signals verified/safe without the aggressive connotation of pure green.

COLOR RATIONALE

Warm parchment (#f9f7f3) → document-like, trustworthy, not a consumer app
Deep ink (#1c1a16) → authority, weight, precision — headers and data
Moss green (#2d5a3d) → verified / matched — calm confidence, not celebratory
Gold (#c8a44a) → partial / warning — needs attention, not panic
Deep red (#9a2020) → unmatched / error — serious, not alarming
Navy blue (#1e4080) → AI suggestion — distinct from human actions

RULE AI suggestions always appear in blue — never in green (verified) or red (error)

Phase 04 — Working Prototype

Full reconciliation interface

Fully interactive. Click any row to open the detail panel. Filter by match status. Use the action buttons (Match, Flag, Reject) to resolve exceptions. Watch the summary stats update as you work.

PAYMENT RECONCILIATION ENGINE · Q4 2024

Clearwater Payments · Daily Run — Nov 18, 2024

Total Transactions

1,284

$48.2M total value

Matched

1,196

93.1% · auto-confirmed

Partial / AI Suggested

$1.1M · review needed

Unmatched

$280K · action required

FILTER:

All

Matched

Partial

Unmatched

REF / ID	COUNTERPARTY	DATE	AMOUNT	DISCREPANCY	STATUS	AI CONFIDENCE

⟡

Select a transaction
to view match details
and AI suggestions

Phase 05 — Reflection

What the design changed

Critical Design Decisions

D1 Sort by exception severity — not date
D2 AI suggestions always in blue — visually distinct from verified
D3 Split-pane keeps list context during review
D4 Summary stats update live as analyst resolves items
D5 "AI Auto-Resolve" batch button for the 78% solvable cases

Next Iterations