Reversing engagement decline at Fortune 500 retail scale · myDG Wallet
Everybody wanted Wallet fixed. There was nothing to fix — there was something to architect. This is the thesis, the disagreement I lost, and the system that's still absorbing new programs two years later.
Wallet had grown one feature at a time — clipped coupons in one tab, DG Cash balance in another, the barcode buried, Saved Deals stranded on the shopping list. Each addition was reasonable in isolation, incoherent as a whole. Eight features across six screens. Multiple orgs, multiple PMs, nobody owning the surface. Engagement declining double digits with no sign of stabilizing — while the business kept asking "can you just fix Wallet?"
What made it urgent: Cash Back was working. I'd led UX for that integration the year before — DG was the first retailer in the discount channel to join the Ibotta Performance Network, 20,000+ stores. It added a real value layer, and its success surfaced the gap underneath. The feature I'd shipped is what exposed the architecture problem.
My product partner brought the decline data. Customers couldn't answer the most basic question: what do I actually have here?
Wallet was an architecture problem, not a features problem.
No single feature was failing. Six screens were splitting one mental model, and users never got a complete answer to "where's my money?" Engagement drops are often design problems, not user problems.
"The user's mental model is the only spec that matters. A PRD that contradicts it is wrong even if it ships."
Three inputs shaped the thesis: the decline data (fragmentation as root cause, not missing features), user research (people think in terms of "my stuff" — six screens meant six partial answers), and AI-augmented exploration — we surfaced directions in days instead of weeks, which let us pressure-test the architecture before asking anyone to commit roadmap to it.
The working loop: brain-dump the problem space → AI synthesis → rough builds → creative direction against them → Figma craft → handoff. Cost-to-restart was effectively zero, so we explored ten rough directions instead of defending one — then boarded the five worth arguing about, killed four with reasoning, and shipped the survivor.
The honest footnote: this was early 2024 — Airtable AI to synthesize research, Uizard and Galileo to spin layouts, years before those tools matured. The instinct was there before the tooling caught up. But the tooling only gave us volume, not judgment.AI generated the options — it couldn't tell us which one mattered. The card stack came from creative direction applied to that volume. That part was ours, not the model's.
One mental model of money. Rewards, clipped deals, payment methods, and order history in a single system — a virtual card stack, barcode redemption at the register, a direct path from Wallet to checkout.
Two moments served — pre-shopping and in-store, one continuous flow.The barcode was elevated to the primary action.
What made it a system, not a redesign — three things:
Nobody owned Wallet, so alignment had to be built. My product partner and I made a paired argument — her data, my architecture — presented as one case to Product, Engineering, and the Chief Digital Officer.
The signal it was working: teams stopped arguing about where features should go and started arguing about whether features fit the model. Same energy, better question.
A customer's saved deals lived on completely different technologies — coupons, offers, and just-for-you on one, cash back on another, rebates on a third. I pushed to unify them into one wallet. The moment that mattered was in-store: most of our customers shop in the store, and at the register it's all about the wallet. If they thought they'd clipped a coupon and it didn't ring up, they had to hunt across screens.
The call was deferred — and the reasoning was sound: the cost was high and largely unknown, the technologies genuinely different, the capability not there yet.I argued for the customer with evidence, not a louder voice. Then I respected the call.
Eight months later the architecture won on its own evidence, and we unified it. That became the Unified Wallet that stands today — slower than I wanted, more durable because it wasn't forced.
The model tempted us to pull everything in — but payment methods and order history belong to the account, not the wallet. So we surfaced access through a slide-in drawer and kept their home in settings. The architecture drew that line, not politics. That's when I knew the thesis held under pressure.
Then the system started absorbing. Cash Back lived inside it from day one. A year later Rewards plugged in — a re-imagined version of an earlier rebate-style program that never fully matured — without a single new primary screen. One component, three progress variants (dollar fill, item dots, trip counter), because how you earn it is the product. Engineering built once. The vision outlived the project.
Sourcing: figures are from the program's own reporting. Context, not causation — the app grew from 5M to 12M+ MAU across my four years; Wallet served that base.
"The discipline isn't adding features. It's building the system that absorbs them."
What I'd do differently: I'd bring engineering into the exploration earlier. We brought them a thesis to react to instead of a problem to shape. The instrumentation disagreement might have resolved in weeks instead of eight months if the people who owned the data plumbing had been in the room while the card stack was still clay. That changed how I run vision work now — engineering joins at the exploration stage, not the commitment stage.
I diagnosed a broken mental model at twelve-million-MAU scale. I didn't apply a framework.