Agentic customer experience roadmap

The context

Amazon's grocery and Same-Day business was scaling fast, and perishables are unforgiving. A delayed laptop generates a complaint. A spoiled avocado generates silence, then churn. That makes customer experience signals the earliest warning system for operational breakdown.

As Senior Product Manager, Tech, I was embedded across Grocery Cart, Checkout and Same-Day Delivery, working backwards from a customer trust gap: the feedback system gave the grocery expansion almost no structured signal.

The problem: the data existed, the intelligence did not

The feedback infrastructure had not meaningfully evolved in years. Surveys went out, complaints were logged, dashboards existed. It could tell you customers were unhappy, but not why, where, or who owned the fix.

Five disconnected signal streams ran in parallel, and nothing distinguished whether an issue started inside the warehouse (UTR, under the roof) or in transit (OTR, on the road). Teams got abstract dissatisfaction scores with no operational accountability attached.

What I did

1. Mapped the full journey before touching anything

Cart, checkout, fulfillment, last mile and post-delivery feedback, end to end. The diagnosis: abundant data, no intelligence layer connecting it.

2. Reframed the product vision

Modernize capture, unify structured and unstructured signals under one taxonomy, then use AI to synthesize complaints for root-cause attribution. The shift from "fix the survey" to "build an intelligence layer" became the basis of the agentic roadmap.

3. Made the survey order-aware

The old survey asked every customer the same questions. I redesigned it with dynamic branching on order content: freshness and substitution questions for grocery-heavy orders, timing and clarity for mixed carts, no grocery prompts on non-perishables. I prototyped the flows with UX research on UserTesting and wrote the PRD covering branching logic, backend triggers and telemetry. Response rates improved 15%.

4. Added AI complaint synthesis

The best signal was buried in free-text care transcripts. Using internal AI synthesis tooling, I clustered complaints and detected severity and trends across thousands of cases, linking spoiled produce to specific fulfillment centers and melted frozen items to OTR delays. It was the first time the organization had UTR/OTR attribution on customer complaints.

5. Unified it in one dashboard

Severity-weighted, frequency-ranked pain points tied to named operational owners, replacing static CX reporting and cutting complaint-to-action latency from days to near-real-time.

6. Defined the agentic roadmap

I defined conversational AI as the next-generation capture layer, integrating with Rufus so customers could describe issues in their own words and feed the synthesis pipeline directly, replacing static post-delivery prompts with context-aware dialogue.

What made it hard

The organizational build, not the technical one. Getting teams to treat a long-neglected feedback mechanism as a strategic priority during an aggressive scale-up took a reframe: customer trust signals are not a reporting function, they are the earliest indicator of whether the expansion is working.

Why it matters for AI products

Data before AI. I unified the signals under a standard taxonomy before layering AI on top. AI on messy data produces confident wrong answers.

AI where it has leverage. I scoped the AI to the one problem it could solve well, free-text synthesis at scale, rather than everywhere.