← All work

Walmart · 2024

My Assistant + Intelligent Operations

Operational work was spread across systems that never explained themselves.

my-assistant / operations
Where is my truck and what should I move?

Task preview

IntentReschedule afternoon truck
ContextStore 2145 · PM shift
ActionMove 3 tasks · notify 2 leads
ConfirmAdjust

Interface recreated for portfolio presentation to protect confidential product information.

My role

Senior UX Designer

Impact

  • 30% increase in ai engagement

Scope

  • AI interaction model
  • Conversational flows
  • Product design
  • Prototyping
  • Cross-functional alignment

Timeline

2024 — 2025

The challenge

Associates were expected to know which system held the answer, and in what order to visit them.

The opportunity

An assistant that reads operational intent and assembles the workflow instead of pointing at it.

What research changed

  • Intent is rarely a question

    Many inputs were tasks in disguise — a question about a delivery usually meant a plan needed changing.

  • Context beats personalization

    Store, shift and role scoped the right answer more reliably than usage history.

  • Generative output needs edges

    Free-form text without structure was read as unreliable, even when it was correct.

Decisions

Decision 01

Generative systems propose; people commit.

  • Let the assistant act autonomously
  • Require confirmation on every action
  • Confirm only high-impact actions

Options considered

What changed my mind
In testing, an action taken without a preview was read as a mistake even when it was correct.
Why
Operational work is accountable work — the person answering for the outcome should authorise it.
Tradeoff
An extra step in every flow, traded for a system people are willing to keep using.

Decision 02

Render answers as structured, typed components.

  • Answer in prose
  • Answer in typed components

Options considered

What changed my mind
Prose forced re-reading; people looked for the same fact in a different place each time.
Why
A predictable shape makes an unpredictable system feel dependable.
Tradeoff
The assistant can express less than a language model could, and each new answer type needs design.

How it changed

  1. V1

    Assistant answers and acts

    The system completed the task it inferred.

  2. User signal

    Autonomy reduced trust

    Associates wanted to see what would change before it changed.

  3. Final

    Propose, preview, confirm

    The assistant proposes an action and renders exactly what will change; the person commits.

Constraints we designed around

  • Business requirement

    Actions touch operational records, so reversibility and clear authorship were non-negotiable.

  • Data availability

    Not every system exposed the context the assistant needed, so the design had to degrade to a clarifying question rather than a guess.

How might we

How might we turn an operational question into a completed task in a single exchange?

Where it landed

Task previews

The proposed action shows exactly what will change before it happens.

Contextual grounding

Store, role and shift quietly scope every response.

Recoverable steps

Generated actions can be reversed rather than escalated.

my-assistant / operations
Where is my truck and what should I move?

Task preview

IntentReschedule afternoon truck
ContextStore 2145 · PM shift
ActionMove 3 tasks · notify 2 leads
ConfirmAdjust
Interface recreated for portfolio presentation to protect confidential product information.

Outcome

30%

Increase in AI engagement

What I took from it

Engagement rose when we removed capability. Constraint is what made the assistant feel dependable.

Next project

CVS Health

Making pharmacy experiences easier