I helped design an AI chatbot that stopped pushing people toward Sales and started actually helping them.

Salesforce

|

4 Months

|

8 Designers + Researchers

TL;DR

Salesforce's existing chatbot had one move: "Contact Sales." Every time. For everything. So my team and I built Fin, an AI chatbot that reads the room.

This is Fin!

Fin guides small business owners through product discovery with empathy, context, and a personality that doesn't feel like a pop-up ad.

Before

Redirected to Sales Team incase of unclear entry

After

Clarifying question asking to be more elaborate

My Impact

I led the personality sprint, defined how Fin speaks and why, and co-designed 8 end-to-end features.

0+

0+

Collaborative sessions

0K

0K

Projected small-business sessions improved

Projected Impact

0+

0+

Use cases designed

The overarching question

Before we designed anything, we had to agree on what we were actually solving for. The current chatbot wasn't broken, it was just built for Salesforce's goals… not the user's.

How might we design a human-centered AI chatbot experience for small business owners — one that helps them find the right product, without making them feel sold to?

We started with the existing Einstein AI chatbot. After a few minutes with it, three things became obvious:

Pushing to Contact Sales

Poor Product Discovery

Being Too Robotic

None of these were accidents. They were consequences of building a chatbot that optimized for conversion, not comprehension.

The Solution

Here's what we built. It wasn't just "make the chatbot nicer."
We redesigned how it thinks, when it talks, and who it feels like.

1

Personality —

This was my sprint to lead. Before we picked a font or a name, I had a question no one had answered yet: What kind of person should Fin be?

The goal was to improve user engagement with the chatbot through tone and language, increase user retention, and design a seamless bot to human transition

2

Comparing Products —

Small business owners come to Salesforce with a mess of options and no idea where to start. The chatbot was making it worse.

Pain Point

Challenging for users to compare and retrieve product details in the chatbot conversation

Solution to this problem

Summarize product key features and create comparison table within the chat

Key Insight

Users expect the chatbot to simplify product discovery by providing quick information to help them decide whether to explore further

3

Chat Timeline —

Information without navigation is just noise. Users were getting answers but losing them immediately.

Pain Point

Endless scrolling to find specific information in the chat

Solution to this problem

  • Real-time timelines of the chat

  • Flexibly open or close the timeline

  • Progress indicators while scrolling the chat

  • Titles on timeline are created in real-time

Key Insight

Users want an easier way to reference relevant info such as product suggestions in chat without scrolling endlessly

4

Chatbot Redirection —

The moment a user left the page, the conversation died. We needed to design for the full browsing journey, not just the chat window.

Pain Point

Users worry they will lose the chat upon switching web pages

Solution to this problem

This helps the users do a dynamic search of different products on the website, by leveraging chatbot tools

Key Insight

Users need assurance that their chat and shared content (like product links) are safe and accessible while browsing

5

Behavioral Triggers —

Users kept typing things like "CRM" or "help" and getting back generic responses. One word in. One word out. Not useful.

Clarifying Questions

Suggestive Prompts

Connect to Agent

User Inactivity

Users often input one word entries and expect relevant responses by an AI Chatbot

Before

Redirected to Sales Team incase of unclear entry

After

Clarifying question asking to be more elaborate

A peculiar challenge…

Theres a problem! We needed to observe how users actually interact with a procurement chatbot and test which personality type felt right. But, no chatbot existed that could do both at once.

So I proposed something a little unconventional.

I suggested leveraging ChatGPT to simulate Einstein AI for conducting contextual inquiries

We set it up to mimic three distinct personalities, drawn from the 16 personalities framework. Each participant got a different version of Fin. Same task. Same questions. Wildly different conversations.

Fun/Joy Personality

Formal Personality

Empathetic Personality

We chose GPT over other Agents for one reason: consistency. Multiple participants, multiple devices, multiple industries and Fin showed up the same way every time.

That consistency is the whole point of a chatbot personality.

Affinity Diagram

Once the GPT problem was solved, we wanted to understand what users actually wanted from a business chatbot.

I know… Its a lot to read so here are the key insights we got from this…

Users value a detail oriented chatbot with actionable insights

They want a bot proactive in anticipating needs and be empathetic, inquisitive & clear

No Jargons unless absolutely necessary with a formal tone

Based on these insights, we built out 3 Characters to build Fin

Not archetypes for their own sake — as a way to stress-test what Fin should and shouldn't be.

The decision to give Fin a character wasn't just about warmth. Salesforce has a whole universe of characters in their ecosystem. Building Fin as a character wasn't a creative indulgence — it was a strategic fit.

Curious but Cautious

Highly adaptable

High EQ

You might be wondering why are we doing this in the first place…

It was an opportunity to create not just a character but a joyful connection between the users and the chatbot. Characters are also a big part of the Salesforce branding and it'd only make sense to have a character for the chatbot to easily merge with the ecosystem

Meet Fin!

Curious but cautious. Highly adaptable. High EQ, sociable, and genuinely helpful.

Named for the thing that keeps a dolphin moving and keeps users from getting lost.

This is Fin!

We didnt get here in one go though…

Salesforce pushed back on two things: how Fin communicated during handoffs to human agents, and how it surfaced product comparisons. Fair notes. We rebuilt both.

The Team's Impact

After a long, yet short, project… We achieved a lot!

0+

0+

Contextual Inquiries

0+

0+

Designed Iterations

0K

0K

Projected small-business sessions improved

Projected Impact

But throughout this process, I learnt a lot too!

  • 01

    Eight designers means eight opinions,

    Eight designers means eight opinions, eight instincts, and eight different definitions of "done." I learned that the goal isn't to get everyone to agree — it's to make sure everyone is solving the same problem. Artifacts, framing documents, and explicit assumptions kept us moving faster than any consensus vote.

    02

    Agile is only effective when it’s treated as a learning system

    Sprints were most valuable when we used them to surface what we didn't know — not to prove what we'd shipped. The moment a sprint becomes a delivery mechanism instead of a learning mechanism, you stop improving.

    03

    Constraints are where the interesting stuff lives

    We couldn't recruit enough users. We didn't have a real chatbot to test with. So we built our own research environment using GPT. Every constraint pushed us toward a solution we wouldn't have found otherwise.

    04

    Good problem-solving requires comfort with ambiguity

    There was no playbook for "design the personality of an enterprise AI chatbot." We had to sit with that uncertainty long enough to build our own framework. That's not a skill you develop by avoiding unclear problems — it's one you develop by walking into them anyway.

That’s how I approached this problem.

This is the kind of problem I like… the ones that are more human than they first appear. Here’s another project where I worked through a different kind of complexity…

Know more about me

Connect on LinkedIn!

Go back

My Work

Know Me

LinkedIn