Visually guided AI prediction building
with Einstein Prediction Builder

PRODUCT DESIGN • enterprise application • university collab

Introduction

Timeline

Aug - Dec 2022

Collaborators

4 UX designers, 1 Data scientist
🤗 Special thanks to Salesforce Einstein's Research, Design and Development team for their mentorship throughout the project!

My role
  • I led the student team and corresponded with Salesforce team while proposing UX improvements to Einstein Prediction Builder (EPB).
  • Conducted research, identified key use cases, as well as multiple rounds of feedback and testing. Produced high-fidelity mockups, and delivered final assets to Salesforce alongwith my team.
Impact

Improved the adoption rate(~30%), reduced overall churn and enhanced user experience of a no-code AI prediction tool, Einstein Prediction Builder by Salesforce.

Overview

The Product

What is Einstein Prediction Builder?

Let's say you are an analyst tasked at creating a report about what happens in your business org. In an ideal world, you would create a data visualization based on predictive data. You would write code, and quickly create a machine learning model. What if you don't know how to code? Or have low AI maturity? This is where Einstein Prediction Builder comes in.

It is a low-code tool to create custom predictions and deploy models. Since it’s low code, users don't have to worry about data ELTing(extract, load, transform), wrangling or tuning.

Image provided by Salesforce.

EPB creates custom predictions about what happens in your business without writing any code!

The Challenge

Improving overall user experience

Salesforce team came to us with multiple problem statements. Adoption was low, tasks were left incomplete and models created with Einstein Prediction Builder were often not deployed. We needed to identify how to make it simply serve it's users.

6% drop-off
between the number of orgs who enabled EPB and those who have deployed the prediction.

More than 1/4
of the orgs that have enabled the predictions are not using them.

15% churned users
Started using EPB, and then stopped it.                        

Data provided by Salesforce


How might we guide non-technical users to complete workflows so that they successfully build
AI predictions in Einstein Prediction Builder?

Jump past solution to research 🚀

The Solution

key persona

Lillian
Biz Ops Interpreter (aka Business Analyst) in a startup. Acts as a bridge between business needs and technical capabilities.


Key Goals

Define company's strategic decisions by making trustworthy cross-BU AI insights. Lillian is tasked at creating reports, & dashboards based on these AI predictions.


Pain Points

• Unfamiliarity with Einstein Prediction Builder.
• Low AI & data maturity. Unable to complete all tasks while creating a prediction.
• Finds current predictive model to be time consuming and confusing to use.

Create Context

Lillian, familiarizes herself with EPB through the introductory interface. She is able to see potential outcomes EPB can bring to her work and further, benefit the business.

feature

Walkthroughs take her through the EPB’s canvas and get her up to speed.  

feature
Focus & Progressive Disclosure in UI canvas

Lillian is able to focus on one task at hand with the help of 5-part canvas with clear segmentation of navigation, progress, working area, assistance, and CTAs.

Improvement
Stay in loop with your process

Phase statuses (icons) on the navigation tree keep Lillian in loop of her prediction progress and also grasp her attention of any errors along the way.

Improvement
Promoting User Guidance

Einstein (Salesforce's mascot) visually guides Lillian throughout, by being ever present on the guidance panel.
It helps her with understanding data, results, and terminology.

✨Bonus: Lillian can leverage Salesforce's trailblazer community to post her questions along the way.

✨Key Improvement
Catch errors early with Data Checker

Lillian can easily resolve errors with the help of the data checker. Visual hierarchy and actionable tips help her in real time.

✨Key Improvement

Discover

Research Overview

Understanding EPB's users

With the focus on understanding AI maturity & machine learning levels of non-technical users, my team and I conducted primary research interviews.

Customers mentioned that Einstein Prediction Builder, as it is today, is not serving their needs.

"I had enabled the prediction, but I am not using it anymore."

"I have atleast one prediction enabled and it is used, but irregularly."

"...Feels like a lot of open ended instructions like 'Name your prediction' and everything depends on what data we have."

"Love Einstein is here to help. Wish they guided me at the start."

Understanding workflows

EPB has a long & complicated workflow. Heuristic Evaluation & User testing enabled us to focus on 'Building Prediction' part of the whole flow. Key observations:-

  1. No indication of progress within each phase's steps.
  2. Unclear sequencing & lack of explanation regarding EPB's phases confusing novice users.
  3. Insufficient knowledge bases of datasets & it's characteristics, as well as model predictions.
  4. Users overlook the options listed below the fold.
Understanding the AI prediction tool market

We performed a light competitive analysis to gauge which elements our competition was prioritizing in comparison to us.

EPB had the core components as its competitors, but it's discoverability and explainability was sub-par compared to our competition.

define

Opportunity Areas

insight 1

Users may lack the knowledge necessary to accurately predict a model. Eg, identifying the characteristics of a dataset.

GUIDANCE

Technical nature of AI predictions warrants written or
visual guidance at each step.

insight 2

57% users abandon tasks in the event of a confusion or an error.

GUIDANCE, PREVIEW

Leverage Einstein to help users recover from errors in real-time.

insight 3

Users not aware of expectations of each step. Eg, which part of the dataset is used for train, test and prediction?

FEEDBACK, DRAG & DROP

Progress of each step should be accessible through the interface. Implement drag & drop functionality to boost explainability of AI model predictions.

insight 4

Even though predictions are enabled, they are either used irregularly(29.2%) or not deployed(27.1%).

EDUCATION, GUIDANCE

Improve model quality by training EPB. Educate non-technical users to accurately create segments and subsets of data.

Make

Design Principles

Visual Learning
Making the experience of EPB more scannable and engaging through visuals.

Clarity

  1. Reduce confusion for low AI maturity users
  2. Add ease of use & focus through tooltips, relevant examples.
  3. Moreover, Salesforce advocates for clarity in their Lightning Design system.

Feedback

Expert Critiques in absence of user testing

A bunch of critiques from Salesforce's product, UX, and engineering team later....

  1. Integration with the Lightning Design system to be top-priority.
  2. Removing Drag & Drop, since it has opportunity to disrupt Segment→Train→Test process of EPB.
  3. Technical & bandwidth limitations to adding dynamic UX copy, gifs, and tooltips across EPB.

See Final Designs again↑🎨

Reflect

Takeaways

Lessons
  1. Confidence
    While I feel user research and interaction design is best practiced in a team, this project helped me to asses and validate skills I bring as an individual contributor.
  2. Design for AI
    This project has been a fruitful exercise in understanding AI applications, their potential, and limitations as well as AI maturity levels of customers.
  3. Lesson in presentations
    I learned the power of narrative while collaborating and presenting complex workflows to the Salesforce team.
Next Steps
  1. Explore components from our direct competitors and how they mould predictions in EPB.
  2. Collaborate with Salesforce's product team to user test our proposed improvements & features. Measure adoption rate from enabling predictions to deploying them.
Up Next...