Catering support, in one place.
Catering brings Whole Foods Market $306 million a year in the first 150 stores alone, and its customer support was spread across disconnected tools, so agents often had to hunt for an order and kitchen teams didn’t always receive the full request. I brought that work into the Amazon Customer Care Console (AC3) so agents could handle catering orders in the console they already used.
My goal was to give agents the order details, customer context and handoff tools they need in one place.
Following an order through the existing process
Before I designed anything, I traced a catering order from the customer to support and then to the kitchen. I used workshops, journey maps, interviews and a tool audit to see where the process broke down.
01Discover
Getting both teams in the room.
I coordinated an interactive catering ecosystem workshop with the Amazon and Whole Foods Catering teams. We listed the questions people ran into during a catering order, grouped them by theme, and used those questions to guide the research.


01Discover
Mapping the work before the software.
I mapped the order journey first, without tying it to any tool, and then added the systems people used at each step. That separated a broken process, like changes reaching the kitchen after prep started, from a tool getting in the way, like agents searching several systems for one order.


01Discover
Listening to the people doing the work.
Support agents and kitchen teams, on opposite ends of the same handoff, described the same problems: missing order details, last-minute changes and too much time spent piecing together what had happened. That gave me three things to solve for. Order details had to be right the first time, changes had to reach the kitchen before prep started, and agents needed one place to understand and resolve an issue.
It is such a waste of time to have to switch between tools or screens to find orders.Support agent
I just want all the actions I need in one place.Support agent
We often receive incomplete or unclear requests, because agents don’t have the right information upfront.Kitchen team member
I always have to search using an email or phone number, and sometimes that isn’t enough to locate the order.Support agent
We don’t always know which orders are urgent or require immediate attention.Kitchen team member
It takes too long to find the right order when a customer isn’t authenticated.Support agent
02Define
Affinity mapping the interviews.
I sorted the interview notes by theme with affinity mapping. Agents typed unique identifiers by hand, repeated steps and escalated simple issues because the order card lacked context, and newer agents learned by trial and error. The kitchen’s side, summarized here, came down to hearing about problems late, working around the system and having no central way to track an escalation.
We relied on back-and-forth calls with customer service just to work out what went wrong. It felt like we were constantly playing catch-up instead of focusing on the food.
This representative quote combines themes raised across the interviews.
- Order details that are right the first time
- Changes that reach the kitchen before prep starts
- One place to check, not four
- Orders arriving incomplete or late
- Last-minute changes heard about second-hand
- Systems that do not talk to each other
- Peak catering season, when none of it can wait
02Define
Seeing the whole tool landscape.
I mapped the tools around the catering journey with developers, project managers and end users. Each touchpoint is the tool used for one step in one sales channel, and across six channels and seventeen steps I found 102 touchpoints: 15 repeated the same information, 18 were unclear or missing, and only 6 worked cleanly from end to end. The direction was clear: stop making agents move between tools to get the full picture.
Duplicated 15 Gap or unknown 18 In place 53 Works well 6 Not applicable 10
Scroll the map sideways to read it.
The map is a table of six sales channels against seventeen journey steps. Each cell names the kind of tool used and its state: in place, works well, duplicated, a gap or unknown, or not applicable. Totals for each state are given in the key below the map.
02Define
Following the handoffs.
The journey map showed the customer’s path, and a service blueprint showed what happened behind the scenes. Two handoffs kept causing trouble, agent to lead and agent to the next team, and those became the basis for the ask-a-lead and warm transfer flows.
02Define
Finding the two paths to the order card.
Agents reach an order card in one of two ways. A signed-in customer selects Need Help on Amazon.com and reaches an agent who can already see the order. Everyone else has to be found through Research and Resolve by order ID, email address, phone number or name, so that unauthenticated path is where I focused.
Usability testing on the existing Research call to action showed agents were unsure how to reach the order card, repeated steps and typed unique identifiers by hand. Only 12 agents had access to it, so the path also had to work for a much larger group.
02Define
Writing the requirements as user stories.
The project goal set what agents had to do for a Catering on Amazon order: create or cancel it, add or remove items, change a quantity, switch between pickup and delivery, change the pickup location, date or time, revise or reprocess a payment and process a refund.
I wrote those needs as user stories from the agent’s side. Every change or cancellation sends the customer an email specifying it and alerts the in-store Team Members. The stories also set every order detail an agent needs, from the customer’s contact details and pickup time to each item’s price look-up (PLU) code or Amazon Standard Identification Number (ASIN), the quantity ordered against the quantity filled, and the full payment breakdown.
03Develop
Reusing what already works.
Before making new components, I reviewed Amazon’s existing support patterns against the catering tasks, and most of the flow could use what agents already knew. That kept the experience consistent between Whole Foods and Amazon, let engineering build on proven components and saved new design for the real gaps.
Scroll sideways to follow the flow.
03Develop
Adding a Catering Card to the existing workflow.
The AC3 workflow already ran in four steps: open Research, select the Research call to action, enter a unique identifier and land on the order card. A separate catering workspace would have meant rebuilding components that don’t carry over, so I added a Catering Card to the existing workflow instead.
I started from the existing order card, took out what catering didn’t need and added back what it did. The unique identifier inputs became one search field that accepts an email address, phone number or customer name. I also asked engineering to build on the AC3 functional workflow, whose back button keeps agents from losing unsaved actions.
03Develop
Filling the gaps in the existing console.
Catering needed a few things the console didn’t already have, including pickup details and a clearer order history view. I designed those pieces to fit the existing Customer Care experience, used Amazon Ember, checked contrast, and made sure status was never communicated by color alone.
Scroll sideways to follow the flow.
03Develop
Ordering the solve card around customer questions.
Feedback on the low fidelity solve card mockups set three rules for the next revision. Every piece of information had to answer a common customer question, the expanders were ordered by how often customers ask each question, and acronyms were written out or explained.
03Develop
Ranking the action stripe by usage.
Review feedback on the action stripe set four rules. Actions are ranked by usage, which leaves room to add more later. Buttons stay visible instead of being disabled, and selecting an inactive one explains why, so the agent can help the customer fix the issue. Actions that can’t apply to an order’s status, like changes to a cancelled order, are hidden.
03Develop
Scoping order edits for the next milestone.
The next milestone added order edits: adding an item, changing a quantity, removing an item and updating the pickup slot. I raised the open questions with the product team, including whether an agent could add several items at once and whether the quantity dropdown should cap at what’s in stock. If only 4 trays of meatballs are left, 5 and up would be greyed out or hidden. Pickup slot changes are only allowed until 48 hours before pickup.
03Develop
Adding to the system with care.
When a component could be useful beyond catering, I added it to the shared system and documented how to use it, and catering-only pieces stayed with the feature. That gave the Amazon team a clear handoff without turning a focused project into a parallel design system.
04Deliver
Testing the work with the people who’d use it.
I took the core flows through visual design and usability testing with agents and kitchen teams, then piloted the updated Research and Resolve workflow with the 12 agents who already had access. Their feedback stayed tied to finding an order, making a change and handing work to the right person, and the final refinement added tooltips and visual aids for new agents.
One console, built for the way agents work
Catering support now sits inside Amazon Customer Care, with the order, the customer’s context and the next actions in one workspace. It soft launched to 150 stores in 2023 and reached every store by the end of 2024. Each flow below is paired with the interview finding it answers.
-
Starting a shift without opening multiple tools It is such a waste of time to have to switch between tools or screens to find orders.
Support agentAgents used to begin by searching across separate systems. Now they choose a region and work type and enter the console with the right context already available. -
Finding an order when the customer can’t authenticate It takes too long to find the right order when a customer isn’t authenticated.
Support agentSometimes an agent has only an email address or phone number, so I designed the search to start from those details, with three guided questions about the order that narrow the results. Once the agent finds the right order, it stays connected to the contact, so they never have to start over in another tool. -
A chat, end to end I just want all the actions I need in one place.
Support agentTranslation is doing the work in this contact, so the agent needs to know which words are theirs and which are the system’s. Every message carries a translated label. The match cards populate with the order already attached, the transcript of five prior messages expands in place, and when one translation fails the error itself offers the transfer, so the agent never leaves to go find a bilingual associate. The contact closes into after contact work in the same panel. -
Asking a lead without leaving the customer We don’t always know which orders are urgent or require immediate attention.
Kitchen team memberKitchen teams said they could not tell which orders were urgent, so an escalation carries that signal in structured fields rather than a free-text note. A platter is missing with the event under an hour away. The request takes an issue type and a short summary, the panel header switches from waiting to the lead's name once they join, and the panel minimizes so the customer keeps being answered while the agent waits. -
Handing a call over warm We often receive incomplete or unclear requests, because agents don’t have the right information upfront.
Kitchen team memberKitchens were receiving requests with details missing, so a transfer sends the issue details along with it. Here a cheesecake has arrived damaged before a dinner. The transfer drawer opens under the session controls and each control only enables once the one above it is set, which makes a half-specified transfer impossible to send. The customer is held while the call connects, and after call work opens with the details already on their way. -
Email in one panel I always have to search using an email or phone number, and sometimes that isn’t enough to locate the order.
Support agentEmail had been living in its own tool, so this flow puts it in the same panel as voice and chat with the same match cards beside it. Four attachments expand and download together with progress on each file, and the order ID copies in one click. Two unresolved emails from the same customer open for context and collapse again. Reply opens the composer in place. -
The same panel, failing18 touchpoints on the ecosystem map were a gap or an unknown.Current-state ecosystem mapI designed the error states at the same time as the happy path, so a failure tells the agent what happened and what to do next instead of leaving them to work around it in another tool. Here the connection drops partway through. Customer issue details does not load and recovers on retry, a download fails and every attachment is marked, then a second attempt brings down three of four files. The last one fails through two more retries, the retry link becomes a link to report the problem, and the customer is answered with the attachments that did arrive.
Keeping the work in one place
Agents can find the order, understand what the customer needs and take action without bouncing between systems.
Building on patterns agents already know
The catering experience uses familiar Amazon components wherever possible, and new user interface (UI) was saved for the work that didn’t fit an existing pattern.
Carrying context through a handoff
Ask a lead and warm transfer send the issue type, the order details and a short summary with the request, so the next person has what they need to help without asking the agent or customer to repeat the story.