KART — Case Study · Amaury Brito

KART — from ambiguity to a shipped system.

Auditors were fixing 3,800+ bad transactions a month by hunting across seven tools with tribal knowledge. We turned that reactive workflow into in-product detection, diagnosis, and guided resolution — using AI-assisted prototyping to move from unclear scope to a prioritized MVP in one season — now delivered as the full experience.

ROLE

Product Designer · design pair + PM

USERS

Sales Audit (edit) · Customer Service (view)

TIMELINE

May 2025 → delivered · dev kickoff Sep 2025

DOMAIN

Back-office transaction auditing

No signal ever said “this transaction is wrong.”

As Kohl’s modernized transactions through Aurus and POM, Customer Service and Sales Audit still couldn’t see, triage, or resolve bad transactions in-flow — missing refunds, wrong tenders, duplicates, return mismatches.

Teams hunted across fragmented tools with manual checks and institutional knowledge. The result: a reactive, high-friction workflow that slowed resolution, eroded consistency, and increased financial exposure.

The project began in ambiguity — even the team struggled to articulate what users needed.

Scope and vocabulary were unclear, and the order schema — Items ↔ Fulfillments ↔ Payments, with partials, returns, and missing keys — was complex and still evolving. We stacked evidence until there was a single story everyone could act on.

Schema deep-dives with engineering

Learned the pre-existing data model at the relationship level — IDs, timing, partials, returns — and mapped sources (Aurus/POM, KOSA, SBC, WorldPay) with their link keys.

Sequenced user interviews

Sales Audit (editors) first, then Customer Service (view-only) — to hear and see the real work: single-screen tab juggling, slow UI, copy-pasted IDs, personal notes as navigation.

Service blueprint

Mapped store balancing and problem identification end to end, so pain points and workarounds landed in one shared picture instead of scattered anecdotes.

Problem statements as a living list

Drafted and refined problem statements throughout — including alignment on the hardest question: which fields are safe to edit, and when.

WHAT THE INTERVIEWS SURFACED · PAIN → WORKAROUND

High correction volume → batch manual fixes, tracked in Excel

Missing or delayed store data → pause work, re-run checks, chase stores

Tender mismatches → tab-juggle four systems to reconcile lines

No real-time validation → find issues after the fact, redo work

AI-assisted prototyping

Research made the AI useful — not the other way around.

We fed Figma Make the discovery outputs as structured facts: the order schema, roles and permissions, and success criteria. Grounded that way, it produced end-to-end flows — Order Inquiry, Order Health & Fix, bad-return edits — with tables and microcopy in hours, not weeks.

Feedback sessions with Sales Audit fed straight back into the prototype — copy, layout variants, and the edit-whitelist updated the same day. The clickable prototype became the alignment tool that ended the ambiguity.

1 · Feed it facts

Schema, roles/permissions, success criteria — discovery as the prompt

2 · Generate flows in hours

End-to-end flows, tables, microcopy, quick layout and detail variants

3 · Iterate with users, same day

Sales Audit feedback sessions; pain points folded back into the prototype

4 · Align stakeholders on the clickable thing

A shared, testable prototype replaced abstract debate — and set the MVP

From hunting to guided fixing — detection, diagnosis, resolution in one tool.

Order Health & Fix panel

A single order-level status with clear reason codes — missing capture, partial-refund mismatch, delayed reversal, orphaned tender, duplicate return — with drill-downs to items, tenders, and returns.

Auto-flagging & prioritization

Rules detect anomalies and rank them by severity and age, so auditors know what to tackle first instead of discovering problems after the fact.

Guided, safe resolution

For flagged returns only: limited, whitelisted inline edits with guardrails, live validation, and a full audit trail. Customer Service stays read-only by design.

Order Health — anomalies auto-flagged, prioritized by state, with expected/actual variance in view

Order Details Review — active issues surfaced inline, with guarded tender edits and audit trail

Because most auditors work on a single screen, the experience was rebuilt around progressive disclosure: drill into items, tenders, and returns without losing your place or triggering full page reloads — review, fix, and confirm on one screen.

What I’d tell other designers about AI

“AI works anytime — but it will never understand your users as you do. Research is what makes the output worth shipping.”

Invest in strong prompt creation: objectives, constraints, data fields, roles, success criteria

Bind AI output to an existing design library so it lands on-brand, on-token

Keep a grounding loop with Figma for components, naming, and tokens

Do the research first — clarity in, quality out

Comfortable with the uncomfortable.

This project started with no clear scope, no shared vocabulary, and a data model even the team found hard to explain. The discipline was the same as ever — evidence before opinions — but AI changed the tempo: discovery fed the prototype, the prototype fed the conversations, and the conversations set a prioritized MVP that engineering could start building on time. The full experience has since shipped.

Ambiguity isn’t a blocker. It’s the raw material of the job.