Andrea Barros — 1-800-Flowers Alexa Skill

Client: 1800Flowers
Role: Design manager and lead for conversational design
Launched: Early 2018

In 2017, 35% of Alexa users had activated at least one Alexa Skill. Voice shopping in the US and UK accounted for $2 billion in sales, with forecasts predicting growth to $40 billion by 2022. Recognizing this opportunity, 1800Flowers aimed to create a skill to capture this emerging market. The goal was to enable customers to easily find and send a floral arrangement for any occasion using Amazon Pay's voice-purchase feature.

Overview

I managed conversational design, flow architecture, and the research protocol. I paired with another designer for this challenge. As there was nothing visual to hand off, our output was the engineering specification. With no supporting documentation from Amazon Pay for voice paths, the engineering team and I had to work in tandem, figuring out specifications and limitations as we go.

Scope and methods

Scope. 4 core scenarios architected to command level, ~30 flow nodes, 8 test participants, one utterance library delivered to the client.
Methods. Conversational flow architecture, utterance testing, affinity mapping, emotional and tonal session coding.
Tools. Figma for flow mapping, spreadsheets for utterance documentation, Zoom for simulated voice sessions.

Problem

Voice commerce, an open category at the time, distinguished itself by enabling purchases within a skill, rather than transferring them to customer service representatives or online checkouts. However, the challenge lay in the absence of a visual interface. Notably, this product was one of the pioneers in integrating Amazon Pay ordering through an Alexa skill, revolutionizing conventional methods of transactions.

Challenges

No documentation for Amazon Pay's voice path. The payment flow didn't appear in the public Alexa developer guide, so the constraints governing the entire checkout were unknown before the project started.
No screen meant no fallback. Every edge case, from a misheard word to an incorrect address or an abandoned order, required a linguistic resolution. There was no visual UI to accommodate ambiguity.
Design had to serve as engineering spec. With nothing visual to hand off, the flows themselves became the build documentation.

Fig. 01 Early whiteboarding session with the engineering team to map out voice conversational flows

Research

Eight participants evaluated the utterance set remotely over Zoom, with me acting as Alexa in real time while they responded naturally. Expressions such as "send the same as last time" emerged organically during live device interaction.

What the research showed

Reorder had to be offered, not requested. Testing showed people didn't spontaneously ask for it — it became a core interaction pattern.
Floristry knowledge isn't a given. Participants couldn't picture species and colors from description alone, a structural constraint voice cannot solve. In hindsight, Amazon launched Echo with a screen that would have been an ideal technology to pair with this skill. It was just not available at the time of this project.
Occasion-based recommendation anchors the flow. Leading with "Who and what's the occasion?" set up the rest of the conversation.

The team gathered feedback from user testing in each purchase stage with sticky notes

Fig. 02 How the team gathered feedback from user testing in each purchase stage

Process

01
Direct working sessions with Amazon's team. Revealed the payment flow's branch points and fallback paths and undocumented voice-purchase constraints.
02
Role-playing user testing. Structured the utterance testing so I could role-play as Alexa live. At the same time, participants spoke naturally, capturing real language patterns in a transcript that became the flow architecture.
03
Converted test transcripts into flows. Affinity mapped the testing notes into conversation branches, then specified each branch to the level an engineer could build from—coded sessions by emotional tone to identify where flows felt natural versus forced.

Fig. 03 Flow diagrams handed off to engineering to scope out order intent per buyer persona

Solution

The skill guides users from initial intent to completed purchase entirely by voice, with no visual interface.

The conversation begins with an occasion-based recommendation. Reordering is proactively offered, as user testing showed most people didn’t think to request it. Arrangement type and size are confirmed in order, with price presented upfront. The Amazon Pay integration completes the transaction within the skill, while all address and payment exceptions are managed through dialogue alone.

What was delivered

Conversational flows. Four core scenarios (new user, returning user, browsing, purchase) architected to individual spoken commands.
Amazon Pay constraints mapped. A working understanding of the voice-purchase flow's branch points and fallbacks, established through direct partnership with Amazon.
Utterance library. A structured reference library of words grounded in how real people speak, built to extend. This became the artifact the client continued to use and expand on.

Impact

“Initial consumer response has been good and provided many valuable insights, with an increasing number of new and repeated shoppers making purchases.”
– Amit Shah, CMO, 1800Flowers.com^[1]

The skill shipped as one of the earliest Alexa skills to complete a purchase end-to-end through Amazon Pay. The client used the flows and utterance library as the foundation for the production system.

Subsequent reporting indicated positive user response with increasing repeat purchases. No completion rates or order volumes were instrumented, but the utterance library became a durable asset the client continued to extend independently.

What it delivered

Design documentation that functioned as engineering spec, removing the translation layer between design and build.
A reference library for voice interactions that the client owned and could maintain independently.

What worked

Direct design and engineering collaboration made development efficient. User testing led to more natural utterances and proactive prompts. Occasion-based flows matched real intent.

What I'd do differently

I'd run a second, larger utterance testing round before launch. In voice, the failures live in the long tail. I'd also fix measurement at the start of the engagement — the work shipped, but I can't point to commercial outcomes.

What I learned

Designing without a screen exposed how much visual interfaces let you defer decisions. A layout can hold ambiguity; a sentence cannot. It was a fine balance of crafting detailed responses without losing the audience. We introduced pauses between long sentences to ask the user if they wanted to hear more, as everyone has various levels of knowledge about floristry.