Einstein Chatbot
We turned Salesforce's chatbot from a scripted sales funnel into a product advisor: it asks before answering, notices hesitation, and never forces a handoff to sales.
Einstein wasn't a generative AI chatbot. It just looked like one.
When we started, Salesforce's chatbot could produce simple responses and push users to contact sales, and that was about it. It had no memory or personality and no way to discover products. Four failures kept showing up.
Endless scrolling with no way to revisit earlier parts of the conversation. Users lost their place and gave up.
After most responses, the chatbot pushed users to contact sales before they had enough information.
Generic, repetitive phrases made the experience feel robotic. No distinctive personality or human-like tone.
Product suggestions were vague and generic. There were no links or comparisons, so users could not act on a suggestion.
"How does a little window like a chatbot get people to be excited about this experience?"
SALESFORCE SPONSOR · THE BRIEF IN ONE QUESTIONWe ran two studies: how it should talk, and what it should do.
Personality research and context research ran in parallel: 16 interviews with three chatbot personas on one side, contextual inquiries across ChatGPT, Amazon Rufus, and Einstein on the other.
Voice: 16 interviews, three personas
We tested three personality archetypes (playful, formal, empathetic) across 16 user interviews using ChatGPT. Participants prioritized content relevance over personality in product browsing, but wanted empathy in troubleshooting and opinions during discovery. The right tone isn't one style. It's knowing when to switch.
No single personality won. Playful worked for browsing, formal for structured answers, empathetic for troubleshooting. Users wanted the tone to change with the task.
Context: how people navigate chatbots
We ran contextual inquiries across ChatGPT, Amazon Rufus, and Salesforce Einstein to understand how people navigate chatbot conversations. Five pain points kept surfacing: threads and journeys, prompts, reliability and transparency, personal information handling, and length of entries.
Users struggled to find previously discussed products, couldn't start new threads intuitively, and felt uncertain whether the chatbot understood short entries. They needed a way to navigate the conversation itself.
Every feature either finds the product or adapts to the conversation.
The research gave us three principles, and every feature argument after was settled against them. Success meant users needed the chatbot less over time.
Assist, don't redirect.
The sales handoff appears only after repeated failed attempts, never before.
Adapt the tone to the moment.
Product browsing gets opinions. Troubleshooting gets empathy and brevity.
Make the information findable.
Every response should be locatable, comparable, and saveable without scrolling the entire thread.
Four features came out of it. Click through them.
The conversation gets a map
A timeline alongside the chatbox generates clickable headings for each section of the conversation: user needs, product suggestions, comparisons. Clicking a heading jumps directly to that part. We tested two designs: a dark tab (hard to find without onboarding) and a skeleton timeline with a clock icon. The skeleton won, but users clicked instead of hovered, so hover-to-expand became click-to-expand.
Testing changed all four features.
We put the mid-fidelity wireframes in front of four business professionals. Four findings, four changes.
Hover doesn't work for timelines.
Users' first instinct was to click. Hover-to-expand fired accidentally over and over, so we switched to click-to-expand.
Comparison tables were too dense.
Information was too tightly packed. Users wanted more filter options, customizable categories, and clearer headings.
Button labels confused users.
Labels like 'Message' confused users. Did the chatbot receive a message, or is this a messaging feature? We rewrote every label.
Reassurance needs a threshold.
Some users found unsolicited reassurance unnecessary. The final design triggers it only after sustained backspacing.