Autodesk, 2016
Pioneering self-service AI at Autodesk
Early machine learning work on Autodesk's virtual agent platform that laid the groundwork for the underlying system patterns I use for generative and agentic UI today.
From zero automation to 86% of issues resolved in under four minutes.
Role
Lead designer & researcher
Team
Cross-functional core pod spanning AI engineering, data science, content, and research.
Talk to Autodesk Virtual Agent
Gaining engineering alignment for a 0-1 automated chat support required proving the technical viability of intent routing. I built a functional prototype on IBM Watson/BlueMix to demonstrate how an NLP model could handle high-stress software failures and integrate with existing support workflows to reduce case volumes.
The demo both validated the architecture and secured cross-functional buy-in to build the core branching logic that ultimately scaled Autodesk Virtual Agent to production.
Built with
Interactive chat prototype
View on a larger screen to interact with the AVA demo.
Establishing the Voice
Moving from prototype to production required a deterministic behavioral matrix for the NLP. Instead of relying on subjective voice guidelines, user research defined clear boundaries that map conversational style directly to resolution efficiency.
Professional Authority
Full sentences, no contractions. Reads clear and authoritative; can feel cold at volume.
“I understand you are experiencing difficulty with your subscription. I will look into this matter. One moment please.”
Balanced Helper
Professional but human; contractions where they fit. This one won with most participants.
“I can help with that. Let me pull up your subscription details. This should just take a moment.”
Friendly Guide
Short, warm lines. Fine for light questions; precision work needs a steadier register.
“Sure thing! Let me check on that for you. Hang tight.”
Design Principles
Transparency
To establish trust, the virtual agent immediately identifies itself as an automated system, states its capabilities upfront, and eliminates simulated typing pauses. Because user data shows that friction and abandonment occur when expectations are mismatched, the interface prioritizes immediate transparency over human emulation and ensures a clean and immediate handoff to a human agent when system boundaries are reached.
Interpretation and branching
The interface must accommodate non-linear real-world inputs, resolving shorthand, domain acronyms, and instances where multiple intents are packed into a single message. The system structure maps complex conversational branching to maintain user context across fluid subscription, licensing, and activation journeys, gracefully managing unexpected subject changes without resetting the session flow.
Bounded responses
User research led to strict boundaries for the NLP. In higher-stress situations such as troubleshooting and error handling, customers prioritized “informative” and “direct” interactions. To meet these thresholds, the architecture restricts open-ended dialogue. The system executes short, deterministic turns with concrete next steps, routing cleanly on the first pass to prevent trust-eroding conversational loops.
Establishing global guardrails
To scale this framework across engineering and content teams, qualitative insights were translated into concrete execution rules:
Zero-Chatter Policy
‘Chatty’ attributes and forced humor scored in the deep negative during high-stress troubleshooting.
Typographic Restraint
Emojis and exclamation points eroded perceived technical competence and were banned from the resolution flows.
Clarity > Brevity
Users preferred slightly longer, structured chunks of text over brief, ambiguous answers.
Production AVA on tablet and phone
Conclusion
The primary goal of introducing this new technology at Autodesk was to resolve support issues before they became tickets. By connecting intent classification and conversational routing directly to Autodesk’s backend APIs, the system bypassed typical support wait times to offer users immediate, 24/7 service. It successfully resolved 86% of handled issues in under four minutes. Ultimately, the NLP routing and conversational framework established here became the foundation for Autodesk’s enterprise virtual assistant.