2. Problem Landscape (2017-18)
▪ 1 million cases/year and growing
▪ 2 days average resolution time
▪ 3-10 systems accessed by agents for resolution
▪ $13-200+ cost/case range
▪ Customers lose time, create duplicate cases, agents swamped.
3. Alternative Solutions
Option Pros Cons
Decision Tree Good for issue triaging Structure is often representative of the
back-office organization and not of
customer intent.
Content Wide coverage of customer problems Dependent on search for discovery
Human Help Will get you to the right answer Costly in terms of resources, time,
effort
Forums/Community Other users can help Needs time, active involvement of
resources to be useful
6. Solution
▪ For Autodesk customers, the Autodesk Virtual Agent (AVA) is a customer service option that
provides a seamless blend of digital and human support across all Autodesk help channels.
▪ It aims to reduce customer effort, provide instant access to products and services, and scale
support.
7. Classic AVA (1.0)
▪ Single web page, standalone
monolithic application
▪ Focus on providing easy access
to automated transactional use cases
▪ Three high-level capabilities:
▪ Recognition (IBM Watson)
▪ Decision-making (IBM Watson)
▪ Conversational (Custom)
▪ Beta voice/video support use cases in
partnership with Soul Machines
9. Research: Virtual Agents
▪ Context-awareness & Urgency
▪ “Well, if I had a serious issue where I felt like I really had to speak to someone and then
got the result of speaking to a robot, I'd be a little disappointed.”
▪ Solutions, not instructions
▪ “It's very direct and I like it because instead of giving the instructions, it just gives me the
solution.”
▪ Avatar + Conversation = Positive experience
▪ "It was a nice personal touch; I will say that. I don't know if that image (AVA avatar) is
just some computer simulated image. But it is a more personable experience."
10. Research: Conversation vs Decision Tree
▪ A “conversational interface” suggests to people a more “friendly, comfortable, personable”
engagement as opposed to the more mechanical interaction of clicking through a decision
tree.
▪ Participants want to ask their question early in the flow.
▪ Free-form text entry is perceived as a direct and efficient way to ask specific questions.
▪ Participants modify their style of text input: Human agents get details; virtual agents get
keywords.
▪ “If AVA gets a user quickly and easily to a human agent who solves their problem, […] will
score points in the eyes of the user.”
11. Redesign Goals
▪ Reduce customer effort
▪ Buy
▪ Download
▪ Manage Access
▪ Self-help
▪ “Allow customers to ask a question in their own words”
▪ Natural for customers
▪ Improve routing efficiency using NLP
▪ Increase understanding of customer intent from collected data
12. Modular AVA (v2)
Ideal State:
▪ Reach customers where they are experiencing issues
▪ Powered by a set of Machine Learning capabilities that automate natural language
processing
▪ Independent of the interface
▪ Consumable by other ADSK self-help
▪ Increased observability
13. High-Level Plan
▪ Recognition
▪ Scale recognition of customer intent using multiple Data Science models
▪ Decision-making
▪ Decouple dialog design from application orchestration
▪ Conversational
▪ Support customers across different interfaces and different environments with consistent
solutions
▪ Technical
▪ Build for re-use by other Digital Help teams
14. Roadmap
▪ Modularize front-end using React
▪ Reduce cost of developing & maintaining custom components
▪ Extract existing application logic out of Conversation
▪ Build components that can serve multiple use cases
▪ Build workflows that represent business processes
▪ And assist customers in their journey
▪ Expand presence with integrations to reduce customer effort
▪ Support existing use cases (download) in new destinations (Account)
▪ Improve observability with event driven architecture & instrumentation
15. Platform Capabilities
Natural
Language
input
Orchestration
Layer
Next Best Action
Layer
Skills model
Intent model
Edu model
…ML model
Self service
widget
Content
Recommendation
Option to contact
Via ContactApp
Conversatio
nal workflow
▪ Customer input along with behavioral data as input
▪ Personalized solution to problem as output
19. Updates
▪ Customer retains agency
▪ Escalation always available
▪ No forced path based
on single prediction
20. Customer Effort Reduction
▪ Localized context-specific workflows on other ADSK
properties
▪ Contributed to
5% improvements in score
▪ Benefits newly WFH users
with quicker access to
home use licenses
21. Customer Effort Reduction
▪ Converted existing articles & manual
review process to automated workflow
▪ Simplified process led to improvements in
lead generation & ultimately revenue
22. Early results
▪ Aggregate case creation from education
support page vs. total page visits
▪ Blue line is when AVA was added
24. Evaluative Research
Finding Learning Action Item
In most cases customer provide
multiple entries, iterating on what they
write.
Customers re-iterate their initial query
to see if the solution improves.
Preserve input bar after first input.
Customer mentioned being
‘disappointed’, or not finding content
helpful if there was no self-service
solution.
Self-service solutions will be preferred
by customers if available.
Improve article prediction model
training.
Some customers expected more
interaction from a digital agent, rather
than ‘just’ the recommendation of
articles.
Customers consider ‘search’ as an
unfavorable comparison benchmark.
Increase coverage of walkthroughs
and interactive solutions.
25. Demand Gen
▪ Product discovery via personalized
conversations between prospective
customers and sales.
▪ Reach different customers on the same
platform.
▪ Provide education about services and
products.
26. Future Work
▪ Transition use cases to new Rasa Architecture with an eye on
▪ Scalability
▪ Maintainability
▪ Observability
▪ New use cases
▪ Demand generation
▪ In-product help
▪ Intelligent agent hand-off
▪ Personalization & data-driven experimentation