( Castorama · Hello Casto · 2023 )
An AI assistant that asks before it sells.
(01) Context
Castorama wanted a conversational assistant powered by ChatGPT and trained on its tools and products database, to recommend the product that really fits each customer’s project.
The challenges
- 01How do we train the GPT module?
- 02How do we keep users within scope?
- 03How do we show GPT’s answers to the user?
- 04How do we design the interface?
(02) The principle
At least 3 questions before any recommendation.
We role-played real situations, like choosing a drill, to define conversation starters, shortcut suggestions and the questions the AI asks to understand each need.
- 8 Users tested the MVP
- 3+ Questions asked per conversation
- A/B Tests on the final interface
- Live On castorama.fr
(03) In pictures
How I worked
(04) My role & process
- 01
Explore & benchmark
Tests with ChatGPT and a benchmark of other industries helped us define what the assistant should and shouldn’t do.
- 02
Flows & pain points
User flows, touchpoints and pain points along the shopping journey.
- 03
Wireflows & chat UX
The conversation, its starters and product suggestions, screen by screen.
- 04
Test the MVP
A prototype tested with 8 users. What we learned went straight into the copy and the prompts.
- 05
Train the model
Prompt engineering with the tech team, with many iterations on the training.
- 06
Pair design & A/B
Final interface designed with a UI designer, then A/B tested.
(05) Close-ups
Product suggestions
The AI suggests products that match the user’s criteria inside the conversation, with shortcuts to get started quickly.
MVP user testing
We checked that users understood the start screen, knew they were talking to an AI and realised the scope was tools and products.
(06) Key learnings
- 01
My first AI project
I learned prompt engineering and the ChatGPT sandbox on the job.
- 02
Better test documentation
Audio transcripts, videos and screenshots of every conversation made the analysis much more precise.
- 03
The cost of AI
Technical limits and the real cost of language models have to be part of design decisions.