Designer
Me
Role
Lead Product Designer
Platform
Web Saas | B2B Fintech
Team
1 PM, 2 Eng, 1 QA
Market Context
When mid-market companies scale headcount rapidly, finance teams lose visibility over how money is actually being spent. Existing corporate card platforms gave employees purchasing power, but offered little beyond static monthly limits, no real-time controls, no spend context, and no way to enforce policy without slowing down legitimate work.
Fraud slipped through. Reimbursement cycles stretched to weeks. Finance managers spent hours each month chasing receipts and manually reviewing transactions that should have been caught automatically. The challenge was to build a smarter card platform that enforced policy proactively, without creating friction for the employees it was meant to serve.
Reframed Challenge
Four gaps that made finance teams ineffective
Alerts with no context
Finance managers received Slack pings with a transaction amount and an employee name. No merchant category, no spend history, no guardrail trigger, just "please approve."
Fraud hidden in the noise
Suspicious transactions sat inside the same queue as routine approvals. Nothing prioritised. Nothing escalated automatically. Investigators started their week with a 94-row CSV.
38% of rules manually checked daily
The rules engine was powerful but invisible. Creating a rule required a support ticket. Turning one off required a call. No one except a system admin could see what was active.
0 of 5 Finance managers could use without training
The original platform assumed CFO-level expertise. Onboarding took two sessions. New finance hires regularly escalated to the founding team for help on basic card issuance.
Solution
01
8 finance manager interviews
3 fraud investigators
Competitive audit
02
Jobs-to-be-done mapping
Problem prioritisation
OKR alignment
03
Design sprint
3 concept directions
Internal critique
04
High-fidelity Figma
2 rounds moderated testing
A/B on alert format
05
Phased rollout
Iteration loops
SCREEN 01
Slack message with a number. After: a full investigation brief — confidence score, transaction count, employee risk history, location anomaly, and two clear actions. Average response time dropped from 18 minutes to under 3.
→ Severity-based triage (Critical / Warning / Info)
→ AI confidence score with plain-English explanation
→ One-click actions: freeze, investigate, mark safe
SCREEN 02
Slack message with a number. After: a full investigation brief — confidence score, transaction count, employee risk history, location anomaly, and two clear actions. Average response time dropped from 18 minutes to under 3.
→ Severity-based triage (Critical / Warning / Info)
→ AI confidence score with plain-English explanation
→ One-click actions: freeze, investigate, mark safe
SCREEN 03
Cards weren't visual — they were rows in a table. The redesign made cards spatial and immediate: colour-coded by status, spend bars showing utilisation, one-action freeze. Card issuance went from 9 minutes to under 90 seconds.
→ Visual card metaphor maps to physical card mental model
→ Spend utilisation bar surface at-risk cards instantly
→ Filter chips collapse the whole catalogue by type
01 · INFORMATION ARCHITECTURE
Severity-first, not chronological
Early feedback showed managers scanning the alerts list top-to-bottom, wasting time on low-risk items before seeing critical ones. I challenged the default "newest first" order.
02 · LANGUAGE & LABELLING
Plain English over technical codes
Rules were labelled "MCC:7995 BLOCK" in the legacy system. Research showed 6 of 8 finance managers didn't know what MCC codes were. I replaced all technical identifiers with conversational labels.
03 · MODAL VS. INLINE ACTIONS
Approve/decline inline, not in a modal
Initial prototype used a modal for every approval. User testing showed constant context switching broke flow. The revised design put primary actions directly on the card with confirmation via toast.
04 · AI TRANSPARENCY
Show the confidence score, not just the verdict
PMs pushed for a simple 'Fraud / Not Fraud' binary flag. My rresearch showed investigators trusted the system more when they could see the reasoning. A 42% confidence score changes how you respond.
Faster alert review time
18 min → 4.7 min avg
Fraud detection rate
Up from 87% pre-redesign
Fraud prevented in Q1
First full quarter post-launch
Card issuance time
Down from 9 minutes
The interface is not the product — trust is
Finance managers weren't frustrated by bad UI. They were frustrated by uncertainty. When I reframed every design question as "does this build or break trust?" the decisions became clearer. The confidence score wasn't a nice-to-have, it was the whole point.
Bring engineers to user research, not just findings
Mid-project, I started inviting an engineer to user sessions. Not to spec, just to watch. The conversations shifted — we stopped debating feasibility of ideas and started building better ones earlier. It saved us two weeks of rework on the guardrails toggle.
Define "done" for the designer, not just the feature
I shipped the cards redesign and immediately moved on. Six weeks later, an edge case in the frozen card flow caused real user confusion. I should have written post-ship monitoring into my own definition of done — not just engineering's.
This will hide itself!