Andela AI Assist

Designing a Generative-AI-powered tool that proactively embeds the right information into employee workflows.

Project Details

  • Product Design

    Primary Practice Area
  • 2023

  • Head of UX & Product Design

    My Role

About Andela AI Assist

AI Assist was a proof of concept product built by Andela in 2022. It's an Al-driven, contextual information retrieval tool meant to plug into popular CRM's & help desk software platforms. The tool leverages GenAI & Machine Learning to seamlessly suggest relevant information that helps support agents resolve customer support tickets. AI Assist embeds on top of existing software & uses GenAI to boost customer support agent ticket efficiency by reducing the time spent searching for information.


When I joined Andela, there was an existing POC version of AI Assist for our internal database, which enabled them to successfully run an 8-week pilot program. I decided to refine the POC, turn it into a MVP, and make sure the MVP can scale to more internal platforms, and beyond, to platforms like Salesforce and Zendesk.

The Design Process

For Whom We Designed

Support Agents

Support agents answer dozens of relatively complex questions per day. On average, they spend between 5 and 15 minutes on each question. These agents tend to be highly empathetic & creative, typically coming from backgrounds in retail or education.

Power Agents

Power Agents are typically Senior Support Agents or Managers. Their typical day is spent onboarding new agents, setting up integrations, triaging problems, training, & tracking insights on impact. They also evaluate potential opportunities to fill in gaps across data stores.


Administrators are focused on finding systems that require minimal implementation costs and configuration support. They are interested in understanding the ROI of potential products, by analyzing impact on agent productivity & customer satisfaction.

Key Problems To Solve

How might we reduce load on support agents?

Customer service can be one of a company's largest budget items, especially when they operate on a global scale. AI Assist's solution helps to shrink that cost by improving agent velocity, and in turn, creating happier customers.

How might we ensure platform agnosticism?

Designs needed to fit neatly inside AI Assist's existing partner platforms (Zendesk & Salesforce), while ensuring that it would integrate seamlessly into future partners as well. The more accommodating the design, the more marketable the product.

How might we design for inflexibly small modules?

Customer service software is typically data-dense, so AI Assist had to fit into unusually small spaces, often with fixed heights and widths. However, in order to be truly useful, AI Assist had to pack in lots of information. This balancing act between data density and compact UI was crucial to the project's success.

How might we foster feedback to train an ML model?

Creating a visual system (layouts, icons, & colors) that made it easy for agents to quickly find the most relevant information, without having to exit their workflow. In addition, AI Assist needed to create a lightweight way for the agents to give feedback to help train the machine learning model, so agents could ensure that AI Assist would continue to improve every day.

Assessing The MVP

The POC version of the product was a simple text interface. The product, AI Assist, analyzed new tickets/cases, and then left a simple, private comment (only visible to a support agent) on the ticket with a recommended customer resolution for the agent.

Intelligence In Context

One of AI Assist's core benefits is that it works silently in the background, as agents go about their normal work. To make it seamless for agents to go between their current ticket & AI Assist's suggestions, I placed the AI Assist module near the ticket, and high up on the page.

Finding The Answers

AI Assist's core technology is powered by GPT-3.5, which connects customer tickets and scenarios with possible solutions. The accuracy of those solutions is only as strong as the data that AI Assist can access. The engineering team created a secure API that enables companies to give selected access to massive data stores-historical help center tickets, Google Drive, Dropbox, Github repos, and other private document repositories. By constantly scanning through this corpus of files, AI Assist is able to find correlations between tickets, find answers to esoteric technical problems, and process the semantics of any type of data set-across 100 languages.

Real-Time Relevant Answers

AI Assist feeds Support agents contextually relevant answers to tickets in a prioritized fashion, and gets smarter and smarter with every use, as support agents are given every opportunity to further train the model by giving thumbs-up or thumbs-down feedback to each suggestion.

Designing For Sales Growth

Administrators and power agents are especially interested in understanding and demonstrating return on investment. An efficient admin dashboard features an A/B test that showcases the performance of AI-Assist-supported agents, compared to those not using AI Assist. This built-in validation feeds leaders the data they need to advocate for keeping AI Assist as a partner.

A Fully-Stocked UI Kit

The final UI kit is spun-off from Andela Epic Design System. It's robust and user-friendly, ensuring that the internal team is fully empowered to drive the product's future. With this full UI Kit in hand, the design team is set up for long-term success, continuing to build, ship, and scale products that meet the same high quality bar.


AI Assist operates with 96% accuracy on deflected support tickets.
Net Promoter Score
Within 6 months of launch, NPS score rose dramatically; from below zero to 85.
Time Saved
AI Assist saves the average support agent 30% of their time spent on tickets.
Average Deflected Support Tickets
AI Assist enables organizations to deflect, on average 20%, and up to 60% of their support tickets.