●BACKGROUND

What is Insigmind?

Insigmind is B2B automotive supplier management Saas platform that provides AI-assisted purchasing solutions to car companies.
In an automotive supply chain, we have two key players, the suppliers and the buyers. Buyers often have to physically visit supplier companies, reach out to them, talk with them one at a time, which is a time-consuming process. On the other hand, suppliers also have trouble identifying the target buyers in an efficient way.

Insigmind is that agent that brings buyers and suppliers to the same platform.

●BACKGROUND

Challenges

This project centers around the workflow of a Buyer trying to search and identify the supplier they would like to buy products from.

●DIVERGE

We Assumed There Are Three Search Routes

In this first stage, we have not officially conducted user research for this project yet. Instead, I conducted several interest-gauging meetings with the two founders, who both have worked as buyers for more than 25 years.

From these expert interviews, I understood that there are three mains ways buyers often use to look for suppliers.

●TESTING

What Do Buyers Think About the Three Search Routes?

When testing my initial explorations from stage 1, we realized that buyers don’t follow the three ways to explore suppliers as we assumed.

We decided to design an open-ended search bar and categorize the results instead.

We would allow buyers to search for anything they can think of in this search bar, similar to search engines such as Google. Instead of asking them to categorize their search into the three buckets, we categorize the results instead.

●DIVERGE

The One-Search Structure Doesn’t Set Insigmind Apart

At this stage, we had a series of meetings with product leads to revisit Insigmind’s unique business goals. The current supplier search design doesn’t reflect our core business goal that differentiates Insigmind from other products.

We introduced the idea of category curation.

• Identify what products the supplier specializes in

• Place these suppliers into categories based on their specialties

• When user searched for a product, they would be taken to the category with all suppliers that are top in the industry at the product

●DEVELOP

Insigmind’s Category System

●DIVERGE

This Brought Us to AI Agents

Buyers need an assistant that could pick apart their needs, analyze them, then provide potential solutions

While I was busy meeting with management to find a better product direction that focuses on this business model, the dev team is already building the AI logic that would take care of the process of parsing through the user’s search words to form connection with our categories.

●DEVELOP

Iteration 1 - AI Agent as a Required Part of the Search Flow

Design Decision

• When users click “search”, we taken them through the AI workflow
• AI matches several possible categories that answers to the user’s needs, have then make selection, then move on to result

Problem

• Forces users to go through the AI workflow when AI should be presented more as an assistant, an optional workflow
• No way to do quick concrete searches and immediately get result
• Users cannot return to reach AI analysis once they’ve entered search result page

●DEVELOP

Iteration 2 - AI Agent Accompanies Results

Design Decision

• Present result page directly after searching is complete
• Surface AI workflow as an option to use
• Conversational chatbot for user to put in more needs

Problem

• AI takes up a large part of the screen, deprioritizing the search result
• Still forces users to go through an AI workflow before they can see their search results

The theme of this design is to highlight AI’s function as an assistant that can be accessed when needed, but does not takeover the core workflow.

One of the most important metrics for test for is whether this AI-assisted search flow is able to generate accurate results that users are satisfied with.

Test Method

We asked 12 buyers to each do 6 searches, 3 concrete searches and 3 vague searches. Based on the result suppliers AI provides, they then provided a scored on how accurate the results were on a scale of 1 to 10.

Results

In the first round, we got a 5.3 score for vague searches, so we started another set of AI-training to raise the accuracy of the model. We do this by asking all our buyer partners to submit their needs on a daily bases to compile a database, which we used to train the model.
We were able to reach a score of 7.8 for vague searches, and 8.9 for concrete searches.

This project focuses on a complex workflow, so I had to be very concise about my documentation and handoff package in order to make sure the developers know how my designs should be implemented. The package consisted of three components

Workflow

WorkflowThis includes a prototype of the happy path. However, during the dev process, I work with developers to identify edge cases and take note of that.

Design System

When I first joined, there was no design or visual, so I had to make a design system from scratch.

Responsive Design

We are designing for desktop, laptops, and tablets, so I identifies key screens that I design four different dimension breakpoints for.

Through this experience, I’ve learned to become a flexible designer, working with a variety of stakeholders, knowing when to push for UX maturity. Most importantly, it took me out of the design-only mindset and taught me how to both lead a project and be a team player.