First Generative-AI Support Assistant in AutoCAD

A first of its kind Gen-AI-powered assistant that is available to 20 million monthly active users to resolve product issues from within their product. Beta launch achieved 80% resolution via self-service, without escalating to agents.

ROLE

DESIGN LEAD

AREAS

STRATEGY, RESEARCH, INTERACTION DESIGN, CONVERSATION DESIGN

PLATFORM

DESKTOP PRODUCT

DURATION

15 MONTHS

Problem

Autodesk’s flagship product, AutoCAD, is a 2D and 3D computer-aided design (CAD) software that has been around since the 80’s and is trusted and used by millions across the world. When troubleshooting an issue in the product, users had to leave the product and waste valuable time browsing various sites for support, hunt for ways to ask their question or contact a human.

Goal

Provide AutoCAD users with a seamless way to find answers to their product-related questions or connect with support, all in one convenient location, enabling them to get back to work faster.

Teamwork

This project was the first of its kind where many different organizations worked together to align and implement this product over the span of 2 years.

🔆 I collaborated with 40+ Business Stakeholders, Sales Specialists, Support Agents, Product Managers, Engineers, Data Scientists, Taxonomists, Data Analysts, Content Strategists, Researchers, Visual Designers, Program Managers, and Scrum Masters.

Approach

STEP 1: Update Universal Help module to a chat

(an extensible, modular help platform), that used a form-based UI and entity-based search to recommend relevant articles for a user’s issue.

STEP 2: Beta in AutoCAD

Update the module to be more chat-like and present top links for their product troubleshooting questions and connect to live chat agents.

STEP 3: Expand to AutoCAD Verticals:

Conduct user testing and monitor interaction patterns to tweak and improve the bot’s responses and user experience over time as users adopted it within AutoCAD and other related products.

STEP 4: Update to Generative AI:

The Beta version was successful in that it helped us learn and tweak the experience and accuracy of its search results, but the engagement was low as the bot was hard to find. So I partnered with the lead Researcher and Product Manager to craft a vision that convinced AutoCAD team to embed the chatbot with the product canvas thus making it available at their fingertips. In early 2024 we later rebranded the bot as the Autodesk Assistant and The last iteration of this evolution I worked on involved evolving it to new RAG technologies to develop the first Generative Q&A chatbot that speaks “AutoCAD”.

MVP

The very first application was deployed on 5 product pages in the US website, delivering guided support for issues with buying, installing and troubleshooting Autodesk products.

The experience was divided into 3 parts:

  1. Issue classification: Users navigated pre-determined decision trees or formulated their own questions

  2. Self-help solutions: curated links to articles, short answers, dynamic links if question was captured, and Watson-based workflows

  3. Human help: variety of options for customers to choose from to connect with an agent, such as phone, chat, case, etc

Issue Classification

Users narrowed down their issue through 2-4 levels of topic selection. I worked with data science and content strategy teams to design several versions of this tree based on case data. Then I partnered with our lead researcher to test labeling and presentation with customers to ensure that topic selection was meaningful and efficient

🔆 Customers prefer typing in their question so we introduced machine learning in a few branches to classify and route issues.

Self-Help Solutions

For branches with topic selection, 2-6 curated knowledge articles or a short answer were presented in the help tool. I worked extensively with content strategy to identify which solutions would be shown and provided guidance on styling and layout of the short answers.

In branches where users entered a question, the recommendations were dynamic and depended on how well the users described their question. I worked with the search platform team to understand their technology enough to offer guidance on structuring better queries and helped fine tune how the results are displayed. We also piloted a self-service workflow where a virtual agent guided them to get a download.

🔆 Clear navigation ensured that users know where they are and get to where they want to go.

There’s always a way out

Users should never feel trapped and always have the option to restart or exit.

🔆 Customers always had an escape hatch to change course or get to a human.

Variety of Human Contact Options

While users prefer self-service, there are times when they face urgent or complex issues where they prefer connecting with a human.

🔆 Utilize customer account information to offer product or subscription based human help options.

🔆 Seamless handoff to live chat, callback, scheduled calls or email case options within the module.

User Research

I worked closely with the Research team to formulate and run several studies to evaluate various aspects of the experience. We used Treejack and First click studies in testing labeling and order of decision tree topics, and ran usability tests to test interactions and flow.

🔆 Initial selection is extremely important so having fewer topics upfront and displaying a short description helps them choose the right path.

Results

  • While engagement with the ““?” icon is consistent month over month, it is overall lower than expected < 1% website visitors

  • 75% of customers who engage with the app click on at least one article or contact an agent

  • Visits to dotcom that had interaction with the help module resulted in 5% higher cart additions and almost 2% higher number of orders placed

“I really like how this is laid out… live chat, schedule a call, create a case….because sometimes you have to go through menu after menu after menu and it just drives you crazy to do it. This is actually very nice.”

— Customer with a Standard Subscription

Post-MVP: Broader - Deeper - Smarter

After the first 6 months, the modular, extensible Universal Help platform was expanded to cover more use cases in more languages across multiple web properties.

  • Global Coverage: Universal Help was gradually deployed to all global sites (>1000 pages) — adapted to 30+ countries and 20+ languages, with geo-specific content. Since support articles and agents were available in only 10 languages, we designed effective ways to guide users from their local language to English and worked with local experts to make cultural and language accommodations.

  • Account Management Portal: The help app was embedded in the customer’s Account and leveraged the authenticated customer and right away guided them personalized help with their downloads, installation and user management activities.

  • In-Product Support: Embedded inside Fusion product canvas — open rate was 41% and it decreased resolution time by 20%;

  • Pop-up nudge: Implemented a time-activated nudge to drive awareness on the site, increasing engagement by 15%

Autodesk Distinction Award for Team Collaboration (2019)

Awarded for exceptional collaboration as Lead Designer for the Universal Help Module, along with Lead Product Manager & Engineering Manager