DynamicsPower! Melbourne AI for everyone: Virtual Agent & AI Builder

AI for Everyone: Dynamics 365 Virtual
Agent & AI Builder
Andre Margono (MVP)
Solutions Architect | Barhead Solutions

Agenda
 Introduction
 Virtual Agent Overview
 AI Builder Overview
 Findings on Virtual Agent and AI Builder
 Design Considerations when Building Chat Bot

Introduction
 AI & Chatbot have become the hot topic
since 2018.
 More and more companies are adapting
and putting their effort in building the AI
and Chatbot offering

The (a bit of) History
of AI/Chatbot in
Microsoft World
 Azure Machine Learning was
made as public preview
approximately 4 years ago (2015)
 Microsoft Bot Framework was
introduced approximately 3 years
ago (2016)
 In September 2016, Satya Nadella
stated Microsoft’s vision to
democratizing AI for every person
 Azure Bot Service was introduced
a year later (2017). The platform
that speeds up Bot development.

Building AI/Bot Capability: 2 years ago
 Bot: you will need a dedicated team of developers (Node.js or C#)
 AI: you will need a team that consist of data scientist, data analyst, data
architect, really mature data mart/warehouse
Original Image src: https://image.slidesharecdn.com/dataiku-workshoptelecomparistech-bnp-november2015-howtobuildasuccessfuldatateam-151120194439-
lva1-app6891/95/how-to-build-successful-data-team-dataiku-11-638.jpg?cb=1448048781

Building AI/Bot Capability: in near future
 AI & Bot: Power Platform team

Introducing the
Dynamics 365 Virtual
Agent
 Currently in Public Preview and only for
US tenant.
 Expected to be GA around October
2019.
 Rapid Bot Development toolkit! Less
hassle in managing build &
deployment.
 Graphical user interface Bot Builder
 Intuitive conversation builder
 Intent prediction – Built-In as “Topic”
 Extensible with Microsoft Flow “Action”
capability
 No code required (Unless
using/building custom Flow connector)

Starts with the “Topics”
 A topic defines the conversation
“Workflow”.
 There are 2 types of topics User Topics and
System Topics
 Out of the box, there are a lot of user topics
for example (good learning material). You
can delete the unnecessary ones.
 Note: System topics can’t be removed, you
can only modify them.

Building the Flow that
Compatibles with Virtual Agent
 The Flows must be created
within the same
environment as the Virtual
Agent
 The Flows must be within a
solution. “My Flow”-only
flows won’t be accepted
 The Flows must be built as
“API” (Receive HTTP Request
as the trigger) and must
accept “POST” request.

Analyse the Bot
Usage and
Performance
 Us the Analytics tab to view
the sessions, feedback and
usage of the bot.
 Useful to learn the
pattern/behaviour when we
want to update the bot
conversation/workflow.

Introducing AI
Builder
 Currently in Public Preview and only for
US tenant. Expected to be GA around
October 2019.
 Comes in 4 main AI Models:
 Binary Classification
 Text Classification
 Object Detection
 Form Processing
 A “Hidden” model (PowerApps canvas
only): Business Card Reader

Binary Classification
 This AI Model is used to predict specific outcome as
Yes/No output based on the trained historical
model.
 Works to predict based on two-options (Yes/No)
field type in CDS + “variable” fields that drive the
outcome.

Form Processing
 This AI Model provides OCR capability for extracting
key–value pair information from the document
 Accepts only PDF, PNG, JPEG (for now)

Object Detection
 This AI Model is used to detect specific object(s) in an
image.
 Can provide the number objects (count) that present in
the image

Text Classification
 This AI Model is able to classify the context of the text
and giving the tag of the text that it reads.
 In short: customisable sentiment analysis

Findings: Virtual Agent
 When building the Microsoft Flow action, always build the response body
directly, building it in variable renders the Flow to be missing/not recognised as
a valid one.
 Array and Boolean output from Flow is not supported yet, but it is mentioned
to be in the immediate roadmap.
 Virtual Agent expects a pre-set JSON response and exact structure is required
for it to work.
 Rich text (images/files) are not supported yet.
 Branding of the bot is not supported yet.

Findings: AI Builder – Binary Classification
 Binary classification will only classify records where prediction = null, where by
default the prediction has default value. A workaround using Flow/Workflow
might be needed to automate the recalculation
 After publishing, the AI Builder classification will run against the record on daily
basis (not real-time).

Findings: AI Builder – Object Detection
 Object Detection: it is not a sophisticated AI model that can classify object with
different properties. E.g: car model classification (SUV, Hatch, Sedan, etc)
 Object Detection: MUST have the category of the object in CDS
 Object Detection: Data feed is a manual feed (no integration to keep pulling
data from a system and automatically train the model)

Chatbot Design Considerations
 Building bot = building apps. Design based on the feature/functionality that
specified as predefined options during the greeting.
 Guide the users by providing prompts as much as possible based on options,
rather than free text.
 Intent/phrase prediction, this is a “prediction”, always put the “exit” clause if the
bot is not “smart” enough predicting the intent.
 Always offer hand-over to human.
 Analyse the usage pattern and refine the bot as it goes along.

Conclusion
 Virtual Agent
 With the Virtual Agent capability, you could build a bot with intuitive graphical interface,
with almost 0 code required
 Virtual Agent + Flow = endless possibility
 AI Builder
 This tool is very useful to build AI Model based on pre-defined ML logic/template
 No code & data science knowledge required to build the AI Model
 Big potential for the future roadmap with more AI Models that relevant to the business

Resources
 https://docs.microsoft.com/en-us/dynamics365/ai/customer-service-virtual-
agent/overview
 https://docs.microsoft.com/en-us/dynamics365/ai/customer-service-virtual-
agent/how-to-flow
 https://docs.microsoft.com/en-us/ai-builder/overview
 Binary classification sample: https://github.com/microsoft/PowerApps-
Samples/blob/master/ai-builder/

DynamicsPower! Melbourne AI for everyone: Virtual Agent & AI Builder

More Related Content

What's hot (20)

Similar to DynamicsPower! Melbourne AI for everyone: Virtual Agent & AI Builder (20)

Recently uploaded (20)

DynamicsPower! Melbourne AI for everyone: Virtual Agent & AI Builder